Track 18 · Advanced

18.1

All AI tools — index

Reference: every tool, side by side — look it up any time

2026-08-08 AI-generated

1Overview

An honest survey of the AI tools your peers actually use — Claude Code, Perplexity, Gemini, Cursor, opencode, Lovable, v0, Base44. Switch between the matrix (task × tool fit), the grid (one tile per tool), and the day-by-day list. Click any tile to read a one-pager: what it does well, where it falls short, what to try in 5 minutes.

In this chapter you’ll explore a complete, side‑by‑side reference of the AI tools that professionals actually rely on—from Claude Code and Perplexity to Gemini and Cursor. You can flip between three views: a matrix that matches tasks to the best‑fit tool, a grid that shows each tool as an individual tile, and a chronological list that tracks daily usage. Clicking any tile opens a concise one‑page summary that tells you what the tool excels at, its limitations, and a quick five‑minute experiment you can try right away.

1.1After this chapter you can

→Pick the right tool for a given research task in under a minute

→Try each tool with a focused 5-minute prompt

→Know which tools pair well together vs. which overlap

1.2How can I compare tool suitability?

You can toggle between a matrix that maps tasks to tool fit, a grid showing one tile per tool, or a day‑by‑day list to see which fits your workflow best.

1.3What details does each tile provide?

Clicking any tile opens a one‑pager that explains what the tool excels at, where it falls short, and suggests a concrete 5‑minute experiment you can try.

2Every tool

A cross-section of what people built browse all recipes →

Base44

Messy spreadsheet of leads — shared live CRM pipeline everyone can update

→ Your whole team works one shared pipeline instead of a messy spreadsheet — with logins, so everyone sees the same live data.

Base44

Shared inbox is a mess — get a live ticket queue with real‑time dashboard

→ Replaces a chaotic shared inbox with a tracked queue and accountability per agent.

Base44

Can’t locate samples or know if they’re expired — get a searchable inventory database

→ No more lost samples or expired reagents discovered mid-experiment; the whole lab queries one searchable database.

I want to…		▶ 3	▶ 2	▶ 8	▶ 3
Write code Python, R, JS — analyze, debug, visualize
Find papers Find scientific sources and cite them correctly
Draft docs Reports, slides, emails, lab notes — fast
Build an app Web tool or lab interface, no coding background needed
Automate workflow Clean data, rename files, script repetitive tasks
Talk through ideas Discuss papers, test ideas, get concepts explained
Properties
Open source / self-host
Data analytics
Templates	●●●	●●	●●●	●●●	●●●
Learning resources

In this course

Dify Open-source LLM app builder with RAG and agent orchestration

Dify

free open source dify.ai →

by LangGenius Inc. · released 2023

Open-source LLM app builder with RAG and agent orchestration

Dify is an open-source platform for building LLM-powered applications — chatbots, document Q&A systems, agent pipelines — through a visual interface. It handles the full stack: prompt management, RAG (retrieval-augmented generation) with a built-in vector store, tool calling, conversation memory, and a publishable chat UI. Where n8n excels at API-level automation, Dify specialises in LLM-centric apps that need knowledge bases: you upload PDFs or web pages, Dify chunks and embeds them, and your chatbot can answer questions grounded in that corpus. For biology students this means: upload 20 papers on a topic, deploy a chatbot, and ask it to synthesise findings across the corpus. Dify supports dozens of model providers (OpenAI, Anthropic, Groq, local Ollama) through a single UI. Self-hostable under the Apache 2.0 licence; cloud version free up to 200 messages/day.

Best for

Build an app
Talk through ideas

Also useful for

Automate workflow
Draft docs

What it does well

One-click RAG: upload a PDF, get a grounded Q&A chatbot in minutes, no code.
Visual workflow editor for multi-step agent pipelines (similar to n8n but LLM-first).
Supports 50+ model providers — swap between GPT-4o, Claude, Groq Llama, and local Ollama without changing the app.
Built-in conversation memory, user session management, and API endpoint for embedding the app elsewhere.
Publishable public URL for sharing a chatbot with collaborators instantly.

Where it falls short

Cloud free tier (200 msg/day) is exhausted quickly in a classroom setting.
Knowledge base retrieval quality depends on chunking strategy — defaults work for most docs but need tuning for dense scientific PDFs.
Self-hosting requires Docker and at least 4GB RAM; non-trivial without a server.

Try it in 5 minutes

Upload a paper and chat with it

Go to dify.ai → Sign up free → Create App → Chatbot → Knowledge → Upload a PDF paper from your field → Deploy → Ask: "What is the main finding of this paper?"

Expected: The chatbot answers with a direct quote from the paper and a source citation.

More prompts to try (1)

Compare two papers

Upload two papers to the same knowledge base. Ask: "What are the key methodological differences between these two studies?"

Pricing

Cloud free tier: 200 LLM messages/day, unlimited RAG queries. Sandbox plan: $0. Professional: $59/mo. Self-hosted: fully free under Apache 2.0.

Open the full Dify course 5 lessons · cheat sheet

⚠️ Watch out

The course Dify instance (dify.32dots.de) is currently paused. For personal use, Dify Cloud gives 200 free messages/day; self-hosting needs Docker.

Groq Blazing-fast LLM inference — Llama and Gemma at 500+ tokens/sec

Groq

free tier groq.com →

by Groq Inc. · released 2024 (public API)

Blazing-fast LLM inference — Llama and Gemma at 500+ tokens/sec

Groq is a hardware + inference company that runs open-weight models (Llama 3, Gemma 2, Mistral, Whisper) on custom LPU (Language Processing Unit) chips at exceptional speed — typically 500-800 tokens per second, compared to 50-100 on GPU-based inference. The speed difference is visceral: a 500-word response appears in under a second instead of five. For biology students the practical implication is fast iteration: testing a prompt, seeing the result, adjusting, and re-running takes seconds rather than minutes. Groq is especially valuable for batch-style pipelines in n8n where you need to process 50+ papers quickly: at 500 tok/s you can summarise 50 abstracts in under two minutes. Free tier available at console.groq.com with generous daily limits for the 8B models; paid tier for heavier use. Important: use your own free API key — the shared course key is restricted.

Best for

Automate workflow
Talk through ideas

Also useful for

Write code
Draft docs

What it does well

500-800 tok/s inference — response latency is 5-10x faster than OpenAI/Anthropic at the same model size.
Free tier is generous for experimentation: 6,000 req/day on Llama 3.1 8B.
Supports Whisper large-v3 for speech-to-text — the voice-to-lab-AI session in this course uses it.
Drop-in OpenAI API compatibility — `base_url: https://api.groq.com/openai/v1`, same format.
Low latency makes real-time applications possible: voice assistants, live annotation tools.

Where it falls short

Model selection is limited to a curated set of open-weight models — no GPT-4o, no Claude.
Rate limits on free tier are per-model, per-minute AND per-day — pipelines processing large corpora will hit daily caps.
Not designed for RAG or knowledge bases — pair with Dify or a vector store for grounded responses.
Free tier does not include fine-tuning or embeddings.

Try it in 5 minutes

Feel the speed difference

1. Get a free key at console.groq.com
2. In the Groq playground, select llama-3.1-8b-instant
3. Ask: "Write a 300-word summary of the central dogma of molecular biology"
4. Note how fast the full response appears vs. ChatGPT free.

Expected: Full 300-word response in under 2 seconds. The speed alone is the demo.

More prompts to try (1)

Voice-to-text transcription

curl -X POST https://api.groq.com/openai/v1/audio/transcriptions -H "Authorization: Bearer $GROQ_API_KEY" -F "file=@your_recording.mp3" -F "model=whisper-large-v3"

Pricing

Free tier: Llama 3.1 8B at 6,000 req/day, 500K tokens/day — sufficient for course work. Llama 3.3 70B: 1,000 req/day free. Paid: $0.05-0.59 per million tokens depending on model. Whisper STT: $0.111/hour of audio.

Open the full Groq course 3 lessons · cheat sheet

⚠️ Watch out

Groq free tier has strict rate limits (30 req/min, 6000 req/day for Llama 3 8B). Do NOT use the 32dots project Groq API key — it was suspended after a silent batch job spent €19.71. Use your own free key.

Cerebras The fastest LLM inference anywhere — open models at 1,800–3,000 tokens/sec

Cerebras

free tier cerebras.ai →

by Cerebras Systems · released 2024 (Inference API)

The fastest LLM inference anywhere — open models at 1,800–3,000 tokens/sec

Cerebras is a hardware + inference company that runs open-weight models (Llama 3.x and Llama 4 Scout/Maverick, Qwen3, DeepSeek-R1 distills, gpt-oss) on its wafer-scale CS-3 system — a single chip the size of a dinner plate (the WSE-3) with thousands of times the memory bandwidth of a GPU. The result is the fastest inference measured anywhere: roughly 1,800–2,600 tokens/sec on Llama models and ~3,000 tok/s on gpt-oss-120B, several times faster than GPU-based providers and even faster than Groq on the larger models. For a biology student the payoff is the same as Groq but more extreme: a long answer is finished before you finish reading the question, and a 50-paper batch job in n8n runs in well under a minute. The free tier is unusually generous — 1,000,000 tokens per day with no credit card — and the API is a drop-in for OpenAI's: change the base_url to https://api.cerebras.ai/v1 and your existing code works unchanged.

Best for

Automate workflow
Talk through ideas

Also useful for

Write code
Draft docs

What it does well

1,800–3,000 tok/s — the fastest inference available, faster than GPU providers and ahead of Groq on larger models.
Very generous free tier: 1M tokens/day with no credit card required.
Drop-in OpenAI API compatibility — `base_url: https://api.cerebras.ai/v1`, same request format.
Runs current open models: Llama 4 Scout/Maverick, Qwen3, DeepSeek-R1 distills, gpt-oss.
Extreme speed makes real-time and agentic loops (many short calls) feel instant.

Where it falls short

Free-tier context is capped (~8K tokens) — not suited to long documents or large RAG contexts without paying.
Open-weight models only — no GPT-4o, no Claude.
Smaller model catalogue than an aggregator like OpenRouter.
Rate limits (~30 req/min on free tier) throttle high-volume batch pipelines.

Try it in 5 minutes

Feel the fastest inference there is

1. Get a free key at cloud.cerebras.ai (1M tokens/day, no card)
2. Open the playground and select llama-3.3-70b
3. Ask: "Write a 400-word explainer on CRISPR-Cas9 gene editing"
4. Watch the full answer appear almost instantly — then try the same prompt on a free ChatGPT tab to compare.

Expected: The full 400-word response streams in roughly a second. The speed is the demo.

More prompts to try (1)

Drop-in OpenAI swap

Take any OpenAI Python snippet and change two lines: base_url="https://api.cerebras.ai/v1" and api_key=$CEREBRAS_API_KEY, model="llama-3.3-70b". Run it — same code, multiples faster.

Pricing

Free tier: 1,000,000 tokens/day, no credit card, ~30 req/min, context capped ~8K tokens. Paid: roughly $0.10–$6 per million tokens depending on model size (e.g. Llama-405B ~$6 in / $12 out).

Open the full Cerebras course 3 lessons · cheat sheet

⚠️ Watch out

Free tier caps context at ~8K tokens and limits you to ~30 req/min — fine for chat and short prompts, but long documents or large RAG contexts need the paid tier. Models are open-weight only — no GPT-4o or Claude.

n8n Visual workflow automation — the lab bench for this course

n8n

free open source n8n.io →

by n8n GmbH (Jan Oberhauser) · released 2019

Visual workflow automation — the lab bench for this course

n8n is an open-source, node-based workflow automation platform that lets you chain APIs, AI models, databases, and files into working pipelines — without writing a backend. Each "node" is a building block: an HTTP request, a code block, an AI call, an IF-branch, a file writer. Chains of nodes become workflows. n8n is the backbone of this entire course: you will connect PubMed APIs, call LLMs, parse XML, route messages via Mattermost, and build multi-stage research pipelines — all inside its visual canvas. Unlike tools that wrap AI features for non-technical users, n8n exposes the full plumbing: you see every API call, every error, every output. That transparency is deliberate — it trains real systems thinking. First released in 2019 by Jan Oberhauser; v1.0 in 2023; fair-code licensed (free self-host, cloud from $20/mo).

Best for

Automate workflow
Build an app

Also useful for

Write code

What it does well

Visual canvas makes pipeline logic readable — you can trace exactly what data flows where.
Built-in AI nodes for OpenAI, Anthropic, Hugging Face — no API glue code needed.
HTTP Request node reaches any REST API (PubMed, CrossRef, Semantic Scholar) with full control over headers, auth, and pagination.
Code node accepts JavaScript or Python for transformations that need more than UI nodes.
Error handling is explicit: retry policies, error branches, and webhook callbacks are first-class citizens.

Where it falls short

Complex branching logic gets visually cluttered quickly — large workflows benefit from sub-workflow decomposition.
Self-hosted version requires a server; cloud version costs money at scale.
Debugging multi-step pipelines takes practice — reading the execution log is a skill in itself.

Try it in 5 minutes

Your first AI workflow in 5 minutes

Open n8n.32dots.de → New Workflow → add a Chat Trigger node → wire it to an AI Agent node → set the model to any available LLM → click "Test Workflow" → type a message.

Expected: The agent replies in the execution panel. You have a working AI pipeline with two nodes.

More prompts to try (1)

Fetch a PubMed abstract

Add an HTTP Request node. URL: https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=34471910&rettype=abstract&retmode=text. Execute it and read the response.

Pricing

Self-hosted (this course): free, unlimited workflows. n8n Cloud starts at $20/mo (starter). The course instance at n8n.32dots.de is free for everyone on the course.

Open the full n8n course 10 lessons · cheat sheet · templates · recipes · downloads

⚠️ Watch out

n8n is self-hosted or cloud-hosted; the course runs a shared instance at n8n.32dots.de. Credentials are shared across the course, so avoid storing personal data in workflows.

OpenRouter One API, every major LLM — including many free models

OpenRouter

free tier openrouter.ai →

by OpenRouter Inc. · released 2023

One API, every major LLM — including many free models

OpenRouter is an API aggregator that gives you a single endpoint and a single API key to access 200+ language models: Claude, GPT-4o, Gemini, Llama 3, Mistral, Hermes, Phi-3, and dozens more — including a substantial catalogue of genuinely free models (google/gemma-2-9b-it, meta-llama/llama-3.1-8b-instruct, mistralai/mistral-7b-instruct, and others). For n8n workflows in this course, OpenRouter is the recommended LLM backend: add a single HTTP Request node, set the Authorization header once, and switch between any model by changing one parameter. This makes it trivial to compare Claude vs. Llama vs. Gemini on the same task, or to build a router that tries a fast cheap model first and falls back to Claude only when needed. OpenRouter also provides unified token usage tracking across all providers in one dashboard.

Best for

Write code
Automate workflow

Also useful for

Talk through ideas
Build an app

What it does well

One API key, 200+ models — switch models by changing one line, no new credentials.
Meaningful free tier: Llama 3.1 8B, Gemma 2 9B, Mistral 7B all free with daily limits.
Ideal for n8n: HTTP Request node + Bearer token + model parameter = any LLM in any workflow.
Unified usage dashboard — see cost per model per day across all providers in one place.
Supports OpenAI-compatible `/chat/completions` format — works with any tool expecting an OpenAI API.

Where it falls short

Free models have daily limits (200-1000 req/day) — fine for experiments, not for production batch jobs.
Adds one hop of latency vs. calling the provider directly — typically 100-300ms extra.
API key leakage risk in shared environments (n8n, classroom notebooks) — treat it like a password.

Try it in 5 minutes

Call a free LLM from n8n

In n8n: add HTTP Request node → Method: POST → URL: https://openrouter.ai/api/v1/chat/completions → Headers: Authorization: Bearer {{your_key}}, Content-Type: application/json → Body: {"model": "meta-llama/llama-3.1-8b-instruct:free", "messages": [{"role": "user", "content": "Summarise the role of p53 in tumour suppression in 3 sentences."}]}

Expected: A clean 3-sentence summary returned in the response body — free, no paid subscription needed.

More prompts to try (1)

Model comparison

Send the same prompt ("Explain mRNA translation in 2 sentences") to three models: meta-llama/llama-3.1-8b-instruct:free, google/gemma-2-9b-it:free, and mistralai/mistral-7b-instruct:free. Compare the outputs.

Pricing

Free-tier: ~10 models available at $0/token (rate-limited). Paid models billed at provider rates + 0% OpenRouter markup. Typical cost: $0-3/month. Minimum top-up: $5.

Open the full OpenRouter course 4 lessons · cheat sheet

⚠️ Watch out

Free-tier models on OpenRouter are rate-limited (typically 10-20 req/min, 200 req/day per model). For production pipelines, budget $1-5/month; do not paste API keys into shared n8n workflows.

LM Studio A polished desktop app to download and chat with local models

LM Studio

free lmstudio.ai →

by LM Studio · released 2023

A polished desktop app to download and chat with local models

LM Studio is a polished desktop app (Mac, Windows, Linux) for running open models locally with no command line. You browse and download models straight from Hugging Face, chat with them in a clean GUI, and — optionally — expose an OpenAI-compatible local server on localhost:1234 for your own code. It handles GPU and memory automatically, so the first run is genuinely beginner-friendly. The app is free, though closed-source. For a biology student it is the easiest way to try several local models side by side — download three, ask each to explain your protocol, and see which is clearest — all on your laptop, nothing sent to the cloud.

Best for

Talk through ideas

Also useful for

Draft docs

What it does well

No command line required — browse, download, and chat with local models from a polished GUI; fully offline and private.
OpenAI- and Anthropic-compatible local API at `http://localhost:1234/v1` — swap one base URL in any existing script or tool.
Side-by-side model comparison: load several models at once and ask each the same question to find the best fit for your task.
Always-on headless mode via the `lms` CLI and standalone daemon — runs as a persistent private endpoint your team or scripts can reach on your LAN.

Where it falls short

Free but closed-source, unlike Ollama or Jan.
Always-on serving needs the `lms` CLI or `llmster` daemon — the GUI app alone is not a background service.
Heavier on resources than a CLI runner; large models need a capable machine.

Try it in 5 minutes

Try models before you commit

In LM Studio, download three different open models and ask each to explain your lab protocol — compare which one is clearest.

Expected: A side-by-side feel for local models with zero setup friction, nothing sent to the cloud.

Pricing

Free to use (the app is closed-source). No subscription; you only need hardware able to run your chosen model.

Open the full LM Studio course 5 lessons · cheat sheet · recipes

⚠️ Watch out

The app is free but closed-source; since LM Studio 0.4+ it can run as an always-on headless server via the `lms` CLI and daemon — not just while the GUI is open.

Chat & research

Claude Desktop All of Claude in a native app — works with your local files

Claude Desktop

free tier claude.com/download →

by Anthropic (Claude team) · released 2024

All of Claude in a native app — works with your local files

Claude Desktop is Anthropic's native desktop application for macOS and Windows (no Linux version as of May 2026) that bundles the Claude chat interface, Claude Code, and the Cowork feature into one place. The key advantage over the web interface is desktop extensions: Claude can directly read and write local files, control your browser via the Claude in Chrome connector, connect to Google Drive, Slack, and Outlook, and monitor code changes as you save them. It is available on all plan types — including the free plan — though Claude Code and higher-model access require paid tiers. For biology students, the most useful daily workflow is dragging a PDF or data file into the chat window and asking Claude to summarize, annotate, or extract tables from it — no upload step, no copying URLs. Memory and project preferences sync across your phone and browser.

Best for

Talk through ideas
Draft docs

Also useful for

Write code
Find papers

What it does well

Drag-and-drop file access: paste a PDF paper or a CSV directly into chat — Claude reads it without needing an upload service.
Desktop extensions connect Claude to your browser, local file system, Google Drive, Slack, and Outlook in one click.
Quick-entry shortcut gives instant access from anywhere on your desktop — useful for quick lookups mid-experiment.
Syncs conversations, memory, and projects across macOS/Windows desktop and the iOS app.
Extended thinking mode available for Pro/Max users: Claude works through complex reasoning problems step-by-step before answering.

Where it falls short

Linux is not supported — Linux users must use the web interface or terminal CLI.
No live web search by default: Claude's knowledge cutoff means it cannot look up a paper published last month without a connector or copy-paste.
Free-tier rate limits are aggressive — after 5–10 messages in a session, you may hit the daily cap and have to wait until the next day.
Claude Code inside the desktop app requires a paid plan, so free-tier users get chat only.

Try it in 5 minutes

Get a paper explained in conversation

Explain the main finding of this paragraph from a CRISPR paper as if talking to a biology bachelor student: [paste paragraph]

Expected: A clear summary with everyday analogies, followed by follow-up questions you can ask.

More prompts to try (2)

Turn rough lab notes into a structured methods section

Here are my rough notes from today's cell culture experiment: [paste notes]. Rewrite these as a Methods section for a lab report, in passive voice, with quantities in SI units.

Extract a data table from a PDF figure description

I have pasted the results section of a paper. Extract all numerical values mentioned for treatment groups A, B, and C and put them in a markdown table with mean ± SD.

Pricing

Free plan included — no payment required to use Claude Desktop for chat. Pro plan at $17/mo (annual) or $20/mo unlocks Claude Code, higher usage, and additional models. Max plans from $100/mo for power-user workloads.

⚠️ Watch out

Free tier has daily usage limits that reset every 24 hours; no live web access by default, so pair with Perplexity for real-time literature searches.

Claude Cowork Delegate whole tasks to Claude — finished work, no terminal

Claude Cowork

paid claude.com/product/cowork →

by Anthropic · released 2026

Delegate whole tasks to Claude — finished work, no terminal

Claude Cowork is the third way to use Claude — a "Tasks" mode inside the Claude Desktop app, next to Chat and Claude Code. It runs the same agentic engine that powers Claude Code, but with no terminal and no coding: you describe an outcome, Claude proposes a plan, and once you approve it works across your local files, folders and apps to hand back a finished deliverable — organised files, a drafted report, a populated spreadsheet, a synthesised literature review. Built for non-technical knowledge workers (research, operations, finance, legal, marketing), it runs shell commands and code in an isolated virtual machine scoped to the folders you authorise, and for bigger jobs coordinates several sub-agents in parallel. Connectors (Gmail, Drive, Chrome, Slack) and job-function plugins add your real tools and know-how, and a /schedule command turns any task into recurring work. For a scientist, it's the assistant that turns a folder of messy data and notes into a clean, finished output.

Best for

Automate workflow
Draft docs

Also useful for

Data analytics
Find papers

What it does well

Hand off whole multi-step tasks, not single questions — describe the outcome and get back a finished deliverable (report, spreadsheet, organised folder).
No terminal, no code: the same agentic engine as Claude Code, in a point-and-review desktop UI built for non-developers.
Works on your real files and apps — reads and writes local folders you authorise and connects to Gmail, Drive, Chrome and Slack.
Coordinates multiple sub-agents in parallel for big jobs, and always shows you its plan before it acts.
Job-function plugins (Finance, Sales, Support, Bio-Research…) bundle the skills, connectors and slash commands for your role.

Where it falls short

Paid plans only (Pro/Max/Team/Enterprise) and desktop-only — macOS and Windows, via the Claude Desktop app.
Agentic file and app access is powerful: keep it scoped to safe folders and use "Ask before acting" for anything sensitive or irreversible.
Scheduled tasks only run while your computer is awake and the Claude Desktop app is open.

Try it in 5 minutes

Turn a messy folder into a clean dataset

In the Cowork tab: "Look through my Downloads/experiment-photos folder, read the data table in each image, and build one spreadsheet with a row per sample (sample id, treatment, value) — then list any images you could not read." Review the plan it proposes, then let it run.

Expected: A finished spreadsheet extracted from your own files, plus a list of anything it could not parse so you can verify the gaps.

More prompts to try (2)

A recurring Friday report

Set up a scheduled task: every Friday at 4pm, pull this week's numbers from my analytics export, drop them into my weekly-report template, and save the result to my Reports folder.

Expense report from receipts

In Cowork: "Read every receipt in my Receipts/September folder and build an expense report grouped by category, with totals and the date range." Review the plan first, then run it.

Pricing

Available on all paid Claude plans — Pro ($17–20/mo), Max ($100–200/mo), Team and Enterprise — through the Claude Desktop app. No free tier; runs on macOS and Windows.

Open the full Claude Cowork course 8 lessons · recipes

⚠️ Watch out

Paid plans only and desktop-only (macOS/Windows). It acts on your real files — keep it scoped to safe folders and use "Ask before acting" for anything sensitive.

Gemini Google's AI — embedded directly in Docs, Sheets, and Gmail

Gemini

free tier gemini.google.com →

by Google DeepMind (Demis Hassabis, CEO; Sergey Brin, core contributor) · released Dec 2023

Google's AI — embedded directly in Docs, Sheets, and Gmail

Gemini is Google DeepMind's flagship large language model, publicly launched December 6, 2023, developed through a merger of Google Brain and DeepMind under CEO Demis Hassabis. Google co-founder Sergey Brin was pulled out of retirement as a core contributor. The model is natively multimodal — it processes text, images, audio, and video in a single prompt — and by 2026 its 2.5 Flash variant with "thinking mode" has become genuinely competitive with Claude and GPT on complex reasoning benchmarks. The practical advantage for students using Google's ecosystem is deep Workspace integration: Gemini drafts full Google Docs from an outline, auto-fills Sheets tables from natural-language descriptions, summarizes Gmail threads, and generates Meet meeting notes automatically. Google AI Plus ($20/month) unlocks Gemini's most capable models; Workspace Business Standard (€13.60/user/month billed annually) bundles AI features across Docs, Sheets, and Gmail.

Best for

Draft docs
Talk through ideas

Also useful for

Find papers
Write code

What it does well

Native Google Workspace integration: generates a full formatted Google Doc, fills in a Sheets table, or drafts a Gmail reply without copy-pasting.
Multimodal by default: upload a microscopy image and ask Gemini to describe the cell morphology — no separate vision model needed.
Long context window (up to 1 million tokens in Gemini 2.5 Pro) — can digest an entire thesis PDF in one prompt.
Google Search grounding: unlike Claude's offline knowledge, Gemini can pull in live search results to supplement its answers.
Free tier is genuinely capable — Gemini 2.5 Flash is competitive for most everyday writing and reasoning tasks.

Where it falls short

Coding assistance lags behind Claude Code and Claude 3.5/4 for complex, multi-file programming tasks — adequate for single-file scripts, weaker on debugging agentic tasks.
Workspace features require a Google account and sometimes a paid plan — not all institutions grant Workspace AI access by default.
Quality can be inconsistent on niche scientific topics vs. Claude or GPT, particularly for less-English literature or specialized biology subfields.
The free tier caps usage without clearly communicating when you're approaching limits.

Try it in 5 minutes

Lab report outline in Google Docs

Create an outline for a 10-page lab report on protein purification by SDS-PAGE. Insert the outline directly into a new Google Doc.

Expected: Full chapter structure with subsections, live in Google Docs.

More prompts to try (2)

Analyze a microscopy image

[Upload a fluorescence microscopy image] Describe the cell morphology, identify any visible structures, and note any anomalies compared to typical HeLa cell staining.

Summarize a long paper

[Upload a PDF of a 40-page review paper] Summarize the key findings in 5 bullet points, then list the 3 papers cited most frequently as foundational to this field.

Pricing

Free tier via gemini.google.com with Gemini 2.5 Flash. Google AI Plus: $20/month (200 GB storage + 2x higher usage). Google AI Pro: $30/month (5 TB + 4x limits). Google AI Ultra: from $50/month (20x limits). Workspace Business Standard includes Gemini at €13.60/user/month (annual billing).

Open the full Gemini course 9 lessons · recipes

⚠️ Watch out

Google Workspace integration is the genuine differentiator; for pure coding tasks or deep literature search, Claude Code or Perplexity consistently outperform it.

Perplexity AI search engine that cites real sources — not just summaries

Perplexity

free tier perplexity.ai →

by Perplexity AI (Aravind Srinivas, CEO) · released Dec 2022

AI search engine that cites real sources — not just summaries

Perplexity AI launched on December 7, 2022, founded by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski — all ex-OpenAI, Google, or Berkeley researchers. By early 2026 the company had reached a $21 billion valuation. It operates as an answer engine rather than a link aggregator: you ask a question in natural language, it searches the live web using its Sonar engine, synthesizes the top sources, and returns a cited answer with clickable references. For biology students the killer use case is literature triage: ask for "three recent papers on CRISPR off-target effects with DOIs" and get a structured, linked summary in seconds rather than manually sifting Google Scholar. Pro subscribers get access to more powerful models (Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro) and can attach files for analysis. The free tier covers most daily search needs with rate-limited model access.

Best for

Find papers

Also useful for

Draft docs
Talk through ideas

What it does well

Live web access with cited sources: every answer links directly to the source URL — you can click through to verify.
Fast literature triage: ask "recent papers on [topic] with DOIs" and get a structured, linked summary in under 10 seconds.
No account required for basic searches — useful as a quick citation-check tool without setup.
Pro mode gives access to Claude, GPT, and Gemini models with web grounding — combining the best of both.
Correctable mid-conversation: if a citation looks wrong, tell it to double-check — it re-searches and often self-corrects.

Where it falls short

Citation accuracy is not guaranteed — the tool can retrieve the right source URL but mis-attribute a specific quote or stat within that paper.
Short-term memory issues: in long conversations it can lose earlier context and repeat itself or go off-topic.
Free tier rate limits have changed multiple times — paying subscribers have been surprised by sudden plan-limit reductions.
Not a replacement for PubMed or Scopus when you need systematic coverage: it samples top results, not all literature.

Try it in 5 minutes

Find 3 recent papers on CRISPR off-target effects

Summarize 3 recent peer-reviewed papers on CRISPR off-target effects with citations and DOI links.

Expected: A numbered list with short summaries and clickable source links.

More prompts to try (2)

Check the current state of a research topic

What is the current scientific consensus on the gut microbiome's role in immune regulation? Cite 2026 reviews where available.

Find a specific protocol

What is the standard protocol for RNA extraction from plant tissue using TRIzol? Cite the original Chomczynski paper and any more recent modifications.

Pricing

Free tier with rate-limited searches using standard models. Pro plan at $20/month (or $200/year) unlocks advanced models, unlimited Pro searches, file uploads, and API access. A student/education discount is available — check perplexity.ai/pro for current terms as pricing has changed frequently.

Open the full Perplexity course 9 lessons · recipes

⚠️ Watch out

Always verify citations against the original paper — Perplexity can occasionally mis-attribute quotes even when the source link is correct.

Claude Claude in your browser — chat, write, analyze, and code

Claude

free tier claude.ai →

by Anthropic · released 2023

Claude in your browser — chat, write, analyze, and code

Claude is Anthropic's assistant in your browser — no install, just a conversation. It chats through ideas, drafts and edits documents, reads PDFs, CSVs and images you upload, searches the web with sources, and writes and runs small bits of code in live Artifacts. It's the lowest-friction way to use Claude: open a tab and go, with a capable free tier and Pro/Max plans for heavier use. For a biology student it's the everyday thinking partner — explain a paper's statistics, turn rough notes into a methods section, or sanity-check an analysis — though it works one turn at a time rather than running long autonomous jobs.

Best for

Talk through ideas
Draft docs

Also useful for

Write code
Data analytics

What it does well

Zero setup — runs in any browser tab, with a genuinely useful free tier.
Strong at writing, analysis, and explaining concepts; reads PDFs, CSVs, and images you upload.
Artifacts run code live in the browser, and web search pulls in current sources.

Where it falls short

Works one turn at a time — no scheduled or long-running autonomous tasks (use Cowork for that).
Web app has limited tool integrations; for MCP connectors or local files use Claude Desktop.
Free tier hits daily caps quickly in a heavy session.

Try it in 5 minutes

Explain a paper's statistics

Paste the methods section of a paper and ask: "Explain which statistical test they used, why, and whether it fits their data — at a level I can repeat to my PI."

Expected: A plain-language explanation with caveats, right in the browser — follow up until it clicks.

Pricing

Free tier with daily limits; Pro $20/mo ($17 annual); Max $100–200/mo for heavier use. No payment needed to start.

Open the full Claude course 7 lessons · recipes

⚠️ Watch out

The web app has limited integrations — for one-click MCP connectors or local-file access, use Claude Desktop instead.

Jan An open-source, offline-first ChatGPT alternative for your desktop

Jan

free open source jan.ai →

by Jan (janhq) · released 2023

An open-source, offline-first ChatGPT alternative for your desktop

Jan is an open-source (Apache-2.0), offline-first desktop assistant — a ChatGPT-style app whose models run 100% locally, so no data leaves your machine. It ships a built-in model Hub for one-click downloads and a familiar chat UI, and it is the most privacy-conscious of the local options: you can work fully offline, then optionally add an OpenAI, Anthropic, or Groq key per thread if you want to call a cloud model from the same app. For a biology student it is the tool for sensitive work — draft notes about unpublished results with the network off — combining privacy, openness, and a friendly interface.

Best for

Talk through ideas

Also useful for

Draft docs

What it does well

100% local and offline-first — data never leaves your machine; Apache-2.0 open source.
Familiar ChatGPT-style UI with a built-in model Hub for one-click downloads.
Can also call cloud models (OpenAI/Anthropic/Groq) per thread if you add a key — one app for local and cloud.

Where it falls short

Local models need a capable machine; smaller hardware limits quality.
Curated Hub is smaller than LM Studio's full Hugging Face browser.
Not built to run as a shared, always-on server.

Try it in 5 minutes

Offline assistant for sensitive data

Install Jan, download a local model, turn the network off, and draft notes about unpublished results.

Expected: A private, ChatGPT-style assistant that works on a plane and never transmits your data.

Pricing

Free and open source (Apache-2.0). Local use costs nothing; optional cloud models cost whatever that provider charges if you add a key.

Open the full Jan course 5 lessons · cheat sheet

⚠️ Watch out

Local models need a capable machine, and the curated Hub is smaller than Ollama's or LM Studio's full Hugging Face access.

Onyx Open-source AI chat + search over your team's apps

Onyx

free tier open source onyx.app →

by Onyx (onyx-dot-app) · released 2023

Open-source AI chat + search over your team's apps

Onyx (onyx-dot-app, MIT open source) is the application layer for LLMs: an all-in-one AI chat and enterprise search that connects to 50+ apps — Google Drive, Slack, GitHub, Confluence, Salesforce and more — and answers with agentic RAG (hybrid search, contextual retrieval and LLM-built knowledge graphs), each claim cited back to the source. On top of search it adds custom AI agents with their own instructions, knowledge and actions; deep research; a sandboxed code interpreter; web search; and Actions/MCP to reach external tools. It works with any model — self-hosted (Ollama, vLLM, LiteLLM) or the frontier (Anthropic, OpenAI, Gemini) — and runs as managed Onyx Cloud or self-hosted on your own Docker/Kubernetes, with SSO and role-based access. Its angle, versus AnythingLLM, is connector-fed, team-scale search across living apps rather than files you upload.

Best for

Find papers
Talk through ideas

Also useful for

Draft docs
Open source / self-host

What it does well

50+ plug-and-play connectors keep living apps (Drive, Slack, GitHub, Confluence, Salesforce…) in sync as one searchable, cited knowledge base.
Agentic RAG — hybrid search + contextual retrieval + LLM-built knowledge graphs — plus custom agents, deep research, code interpreter, web search and Actions/MCP.
MIT open source with self-host on Docker/Kubernetes; SSO and role-based access make it a permissioned, team-scale private ChatGPT.

Where it falls short

Team-scale features (SSO/SAML, granular access controls, white-labeling) are Enterprise or self-hosted, not the free Business trial.
Standard deployment is heavier than a desktop app — vector + keyword indices, workers, Redis and blob storage — so it wants real infrastructure.
Answer quality still tracks your connectors and the model you point it at; garbage in, grounded-but-wrong out.

Try it in 5 minutes

Ask one question across two connected apps

Start an Onyx Cloud trial (or `docker compose up` the CE), connect Google Drive and one more source, then ask a question only answerable by combining them — and open the citations.

Expected: A single cited answer stitched across both sources, with links back to the exact documents.

Pricing

Free and open source (MIT Community Edition) to self-host. Onyx Cloud is a managed SaaS at about $20/user/month (Business, annual) with a free trial; Enterprise adds SSO, on-prem/region deployment and white-labeling at custom pricing.

Open the full Onyx course 7 lessons

⚠️ Watch out

Full Onyx (not the sub-1GB Onyx Lite) runs a vector index, background workers and blob storage — it wants a real server, not a laptop.

AnythingLLM A private, all-in-one ChatGPT for your own documents

AnythingLLM

free open source anythingllm.com →

by Mintplex Labs · released 2023

A private, all-in-one ChatGPT for your own documents

AnythingLLM (by Mintplex Labs) is an all-in-one, private AI workspace: drop in PDFs, documents, or a whole folder and chat with them — with citations — in a familiar ChatGPT-style UI. It is provider-agnostic, so you can run a fully local model (built-in, or via Ollama / LM Studio) or bring your own cloud key (OpenAI, Anthropic, Gemini). When you point it at a local model, nothing leaves your machine. It ships as a one-click desktop app for Mac/Windows/Linux and as a Docker server you can self-host as a shared, always-on team endpoint that exposes an OpenAI-compatible API. Its angle, versus Ollama / LM Studio / Jan, is document-RAG and AI agents bundled in — those tools run models; AnythingLLM turns a model into a private assistant over your own knowledge. Open source under MIT.

Best for

Find papers
Talk through ideas

Also useful for

Draft docs
Data analytics

What it does well

All-in-one private RAG — chat with your own PDFs/docs (with citations), and nothing leaves your machine when you use a local model.
Provider-agnostic: one app for local (built-in / Ollama / LM Studio) and cloud (OpenAI / Anthropic / Gemini) models.
MIT open source, with both a one-click desktop app and a Docker server you can host as a shared, always-on team endpoint.

Where it falls short

Focused on documents/RAG and agents, not raw model management — for browsing and one-click model downloads, Ollama or LM Studio go deeper.
The OpenAI-compatible API and multi-user features need the self-hosted/Docker build, not the desktop app.
Local models still need a capable machine; answer quality tracks your hardware.

Try it in 5 minutes

Chat with your own papers, fully offline

Install AnythingLLM, pick a local model, upload a few PDFs into a workspace, and ask it to summarise and cite them.

Expected: A private ChatGPT that answers from your documents with citations — nothing sent to the cloud.

Pricing

Free and open source (MIT). The desktop app and self-hosted Docker server cost nothing to run; optional cloud LLM providers cost whatever that provider charges if you add a key.

Open the full AnythingLLM course 9 lessons · cheat sheet

⚠️ Watch out

Its OpenAI-compatible API and multi-user server live in the self-hosted Docker build, not the one-click desktop app.

📓 Open Notebook An open-source, privacy-focused alternative to Google's Notebook LM

📓

Open Notebook

free open source www.open-notebook.ai →

by lfnovo (open-source community) · released 2025

An open-source, privacy-focused alternative to Google's Notebook LM

Open Notebook is an open-source, privacy-focused AI workspace that lets you upload PDFs, web pages, audio, and video — then chat with them, generate multi-speaker podcasts, and search intelligently across all your content. It supports 18+ AI providers (OpenAI, Anthropic, Ollama, LM Studio…) so you can run everything locally for free or use cloud models for more power. Unlike Google's Notebook LM, your data never leaves your machine, you control the AI models, and you can generate professional podcasts with 1–4 custom speakers. It ships as a Docker Compose setup (SurrealDB + the app) that you can run on your own machine or a VPS. Open source under the Apache 2.0 license.

Best for

Find papers
Talk through ideas
Draft docs

Also useful for

Automate workflow
Data analytics

What it does well

Multi-modal content — upload PDFs, web pages, audio, and video, then chat with everything together.
Multi-speaker podcast generation — turn your notes into a professional podcast with 1–4 custom speakers.
18+ AI providers — use OpenAI, Anthropic, Ollama, LM Studio, or any OpenAI-compatible endpoint.
Full-text and vector search across all your content — find what you need across your entire notebook library.
Self-hosted and private — your data stays on your machine; run a local model for zero data leakage.

Where it falls short

Requires Docker for self-hosting — not a one-click desktop app like AnythingLLM.
Newer and smaller community than AnythingLLM — fewer pre-built skills and integrations.
Citations are described as "basic" on the project roadmap — not as mature as Google Notebook LM's source linking.
SurrealDB is a new database — you are trusting an unfamiliar data layer for your research.

Try it in 5 minutes

Chat with your papers, fully offline

Install Open Notebook with Docker Compose, connect it to a local Ollama model, upload a few PDFs, and ask it to summarise and cite them.

Expected: A private AI workspace that answers from your documents with citations — nothing sent to the cloud.

Pricing

Free and open source (Apache 2.0). You only pay for the AI model API key you connect — or run everything locally with Ollama for $0. Docker Compose deployment is straightforward: one docker-compose.yml file and you are running.

Open the full Open Notebook course 5 lessons · cheat sheet

⚠️ Watch out

Requires Docker for self-hosting; the free tier is fully functional but you bring your own AI model API key (or run a local model via Ollama).

ChatGPT OpenAI's assistant — broadest features, strong data analysis

ChatGPT

free tier chatgpt.com →

by OpenAI · released 2022

OpenAI's assistant — broadest features, strong data analysis

ChatGPT is OpenAI's assistant — the broadest feature set of the mainstream chatbots and the most mature enterprise tier. It chats, drafts, and explains; its Advanced Data Analysis runs Python on files you upload, making it genuinely strong for stats and plots; and it searches the web with citations when you turn search on. The catch for scientists is data: consumer and Plus chats train the model by default, with a one-click opt-out in Data Controls, and only Team/Enterprise tiers are contractually no-train. For a biology student it is a capable everyday assistant — just keep patient and unpublished data out of the free tier, and click through every cited link.

Best for

Talk through ideas
Data analytics

Also useful for

Draft docs
Write code

What it does well

Advanced Data Analysis runs Python on your uploads — strong for statistics, tables, and plots.
Broadest feature set: voice, image generation, web search with citations, custom GPTs.
Most mature enterprise tier — contractual no-train, SSO, audit logs, EU residency via Azure.

Where it falls short

Consumer/Plus chats train the model by default — you must opt out in Data Controls.
Free tier is not safe for confidential or patient data even after opt-out.
Can fabricate citations — always confirm DOIs and that sources say what it claims.

Try it in 5 minutes

Catch a fabricated citation

After ChatGPT cites papers, ask: "Give me the DOIs and confirm each one actually exists and says what you claimed."

Expected: A reality check on its sources — the habit that keeps AI-assisted literature work honest.

Pricing

Free tier; Plus $20/mo; Team and Enterprise tiers add contractual no-train, SSO, and audit controls. Data residency available via Azure OpenAI.

Open the full ChatGPT course 10 lessons · recipes

⚠️ Watch out

Chats train the model by default — turn off 'Improve the model for everyone' in Data Controls, and use Team/Enterprise for confidential work.

💬 Open WebUI The ChatGPT-style front door your whole team logs into

💬

Open WebUI

free tier openwebui.com →

by Timothy Jaeryang Baek — began Sept 2023 as "Ollama WebUI", renamed Open WebUI in Jan 2024 (Ollama raised a trademark concern); now community-maintained · released 2023

The ChatGPT-style front door your whole team logs into

Open WebUI is the piece that makes a local model feel like a product instead of a curl command: a self-hosted, ChatGPT-style chat app your whole team logs into, with role-based access control and user groups so an admin can decide who sees which models. It talks to Ollama, LocalAI, vLLM, LM Studio, or any OpenAI-compatible endpoint — so it sits on top of whatever engine your lab is running, unchanged. Beyond chat: Channels (shared, real-time spaces where people and models collaborate in one timeline), a Notes workspace, multi-model side-by-side conversations, and — on the Enterprise tier — LDAP/SSO/SCIM provisioning and usage-analytics dashboards. For a lab, it's the difference between 'ask Dave to SSH in and check the model' and everyone just opening a URL.

Best for

Talk through ideas

Also useful for

Automate workflow
Find papers

What it does well

Real multi-user product: logins, RBAC, and per-group model access — not a shared, unauthenticated endpoint
Backend-agnostic — points at Ollama, LocalAI, vLLM, LM Studio, or any OpenAI-compatible API without caring which
Built-in RAG, web search, and a prompt library so non-technical labmates get real value on day one
Channels + Notes turn it into a shared workspace, not just a chat window

Where it falls short

It is a UI, not an engine — you still need something like vLLM, llama.cpp, or LocalAI answering behind it
Enterprise features (SSO/LDAP/SCIM, analytics, white-labeling past 50 users) are a paid, quote-only license
Admin has to actually configure RBAC and groups — a default install is wide open to whoever signs up first

Try it in 5 minutes

Give the whole lab one login page

Run Open WebUI in Docker pointed at your existing Ollama or LocalAI endpoint, register the first account as admin, then invite two labmates and put them in a "read-only model" group.

Expected: Two teammates chatting with your local model from their own browser tab, each with their own history — nobody touching the terminal.

Pricing

Free and open (source-available "Open WebUI License") to self-host at any scale; the license only requires keeping Open WebUI branding visible once a deployment tops 50 end users in a rolling 30-day window. A paid Enterprise License removes that requirement and adds SSO/LDAP/SCIM + support — pricing is quote-only.

⚠️ Watch out

Free and open to self-host at any size, but keeping the "Open WebUI" branding is only optional under 50 users in a rolling 30 days — past that, white-labeling needs a paid Enterprise License (contact sales, no public price list).

🧮 Abacus.ai One bill, 25+ models, plus an agent that builds and one that lives in WhatsApp

🧮

Abacus.ai

paid abacus.ai →

by Abacus.AI, Inc. (Bindu Reddy, CEO & co-founder; San Francisco) · released 2019 (company); ChatLLM 2024; DeepAgent and Claw added 2025–2026

One bill, 25+ models, plus an agent that builds and one that lives in WhatsApp

Abacus.ai (San Francisco, founded 2019 by Bindu Reddy and Arvind Sundararajan) bundles three products behind one login: ChatLLM, a chat surface that routes across GPT, Claude Opus, Gemini, Grok, DeepSeek and 20+ others; Abacus AI Agent ("DeepAgent"), which browses multiple sites, builds and deploys a working web app with a real database, drafts presentations, and connects to Gmail/Google Workspace/Jira; and Claw, an always-on agent that lives inside WhatsApp, Telegram, or Slack, keeps a persistent memory file of your preferences and ongoing work, and runs on a schedule. Everything is metered in a shared credits currency rather than per-product pricing. The company states chats are not used for model training and cites SOC-2 Type II and HIPAA compliance, but publishes no EU-hosting or data-residency option for the consumer tiers.

Best for

Automate workflow
Build an app

Also useful for

Talk through ideas
Draft docs

What it does well

One credits pool spends across 25+ chat models, image generators, and video generators — no per-provider subscription math.
DeepAgent produces a deployed, working app (auth + database included), not just a code snippet — closer to the app-builder category than a chatbot plugin.
Claw is a genuinely persistent agent: it reads a structured memory file before responding and reaches you where you already chat (WhatsApp/Telegram/Slack), not in a separate app.
Unlimited-use models exist even on Basic, so light chat use can avoid the credit meter entirely.

Where it falls short

The credits system is the product's real complexity: video, agent tasks, and "SuperComputer" runtime all draw from the same pool at different, easy-to-misjudge rates.
The Basic tier's 75%-usage throttle is designed to make you upgrade — read it as a Pro-tier upsell, not a technical limit.
No published EU data-residency option; for confidential or regulated data, treat this the same as any other US-hosted consumer SaaS.
Vendor lock-in on the DeepAgent-built apps: they deploy to Abacus's own hosting, not a portable export.

Try it in 5 minutes

Compare three models on the same question in one thread

Ask ChatLLM one research question, then explicitly switch the model dropdown between GPT-5.6, Opus 4.8, and Gemini 3.1 Pro on the same prompt and compare the three answers.

Expected: Three visibly different answers/styles from one login, with the credit cost of each shown per message — the core aggregator trade-off in one screen.

Pricing

Basic $10/mo (first month $7): 20,000 credits, 3 Agent conversations, capped Studio use. Pro $20/mo: 30,000 credits, unrestricted Agent + Studio, SuperComputer access. Enterprise: custom, starts around $5,000. A handful of models (GPT-5 Mini, Gemini Flash, Grok Code Fast, GLM 4.7, Kimi K2, Llama 4) run credit-free.

⚠️ Watch out

A 20,000-credit monthly allowance sounds large until an agent task or a video generation burns through it in minutes — the Basic tier throttles you at 75% usage to make the credits last the month.

🐝 Poe Quora's model router — one points balance spends across 200+ bots

🐝

Poe

paid poe.com →

by Quora, Inc. · released 2023

Quora's model router — one points balance spends across 200+ bots

Poe is Quora's AI aggregator: a single "points" currency you spend chatting with official bots for GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash, Grok, and image/video models like Nano-Banana-Pro and Veo-3.1, plus thousands of user-built "bots" (prompt wrappers and small apps on top of those models). It is the simplest, oldest version of this category — no autonomous agent, no browsing, no scheduled tasks, just model access. Pricing runs from a $4.99/mo Starter tier (10,000 points/day) to a $239.58/mo Pro Max tier (8.25M points/month), plus a free tier with a small daily allowance. Quora, Inc. is the data controller; official first-party bots do not train on your chats or use them beyond producing the response, but third-party developer bots can, so which bot you pick inside Poe matters as much as which platform you pick.

Best for

Talk through ideas

Also useful for

Draft docs

What it does well

The cheapest entry price in this category by a wide margin ($4.99/mo vs $10–25/mo elsewhere).
Largest raw model catalogue (200+ bots) because anyone can publish a bot on top of a base model.
Clear, published stance that official bots do not train on your chats and delete your data on account deletion.
No feature bloat — if you only want model-routing and nothing else, Poe does not push an agent tier on you.

Where it falls short

Third-party bots (not the official OpenAI/Anthropic/Google ones) can use your chats to train their own models — read which bot you're actually talking to.
No autonomous agent, browsing, or scheduling — for that you need Abacus.ai, Genspark, or Perplexity Max instead.
Points consumption varies wildly per bot and per model, making real monthly cost hard to predict in advance.
No EU-hosting or data-residency disclosure found; Quora is the controller for all markets.

Try it in 5 minutes

Chat the same prompt through three official bots

Send one prompt to the GPT-5.5, Claude-Opus-4.7, and Gemini-3.5-Flash official bots in Poe and compare the points each one charged.

Expected: Three answers and three different points deductions in the same sidebar — the cost-per-model trade-off made concrete.

Pricing

Free tier with daily points. Starter ~$4.99/mo (10k points/day). Premium ~$19.99/mo (660k points/month). Premium Plus ~$49.99/mo. Pro ~$99.99/mo. Pro Max ~$239.58/mo (8.25M points/month). Annual billing saves about 17%.

⚠️ Watch out

No agent or automation layer at all — Poe is deliberately just a chat router, so it will not build an app or run a scheduled task for you.

✨ Genspark Super Agent over 9 LLMs and 80+ tools, plus its own persistent "Claw" agent

✨

Genspark

paid genspark.ai →

by Genspark Inc. (affiliate of MainFunc Inc.; entities in Singapore, UK, Mexico) · released 2024; Super Agent + Claw + Workspace 4.0 shipped through 2026

Super Agent over 9 LLMs and 80+ tools, plus its own persistent "Claw" agent

Genspark (Genspark Inc., a MainFunc Inc. affiliate with entities in Singapore, the UK and elsewhere) markets itself as moving from "AI tool" to "AI employee": Super Agent combines nine core LLMs with 80+ specialised tools to write reports, build slide decks, make calls, and generate video, and 2026 added Genspark Claw — its own always-on messaging-app agent, directly parallel to Abacus's Claw. A free tier gives 100 credits/day; paid tiers (Plus $24.99/mo, Pro $249.99/mo) scale credits 100x and add 1TB+ storage. Genspark states a Zero Training and Zero Data Retention policy and cites GDPR compliance and SOC 2 Type II for enterprise clients, though no EU-region hosting option is published for the consumer tiers.

Best for

Automate workflow
Build an app

Also useful for

Talk through ideas

What it does well

Broadest single-product tool sprawl in this category — 80+ built-in tools (phone calls, slide decks, video, Office plugins) beyond just chat.
Genspark Claw mirrors Abacus Claw feature-for-feature (persistent memory file, messaging-app channels, cron jobs) — useful to teach as the same idea from a second vendor.
Explicit Zero Training / Zero Data Retention policy stated in its privacy documentation, stronger wording than most competitors.
AI chat and image generation cost zero credits on paid plans (through Dec 2026), so routine chat use doesn't drain the meter.

Where it falls short

The jump from Plus ($24.99) to Pro ($249.99) is a 10x price step for 12.5x the credits — there is no useful middle tier.
Commercial-use rights on generated content are explicitly time-boxed to Dec 31 2026 in current terms — a real gotcha for any client-facing use.
Corporate structure spans multiple jurisdictions (Singapore/UK/US-linked MainFunc) which makes "who actually processes my data, and where" harder to answer than for a single-country company.
Heaviest actions (phone calls, video) burn credits fast enough that the advertised monthly allowance is easy to blow through.

Try it in 5 minutes

Have Super Agent build a one-page site, then check the credit cost

Ask Genspark Super Agent to research and build a one-page website comparing three lab equipment vendors, then check how many credits the task consumed.

Expected: A working page plus a credit-cost breakdown — makes the "many tools, one meter" trade-off visible in a single task.

Pricing

Free: 100 credits/day, ~1GB storage. Plus: $24.99/mo ($19.99/mo annual), 10,000 credits/month. Pro: $249.99/mo ($199.99/mo annual), 125,000 credits/month, 1TB storage. Team: 12,000 credits/seat/month with centralized billing. Credit top-up packs from $20/10,000 credits, valid 3 months.

⚠️ Watch out

Commercial-use rights for AI-generated content are only guaranteed on paid plans through end of 2026 — a time-boxed guarantee, not a permanent one, so re-check before relying on it for client work past that date.

🔎 Perplexity Max The $200/mo tier that turns Perplexity into a model-council aggregator

🔎

Perplexity Max

paid www.perplexity.ai/hub/pricing →

by Perplexity AI, Inc. · released Jul 2025

The $200/mo tier that turns Perplexity into a model-council aggregator

Perplexity Max is the top-of-line ($200/mo, $2,000/year) subscription of the answer-engine already taught in "Frontier chatbots" (see the `perplexity` entry) — included here because Max is where Perplexity stops being a single-model product and becomes an aggregator: it bundles sonar-pro, gpt-5.4, claude-opus-4.8, gemini-2.5-pro and grok-4, unlimited Pro Search, unlimited Labs, and full Comet access under one price. Independent reporting (not confirmed on Perplexity's own pricing page, which returned an access error when fetched directly this session) describes a "Model Council" feature that dispatches a query across several of these models and synthesizes one answer, plus autonomous, scheduled "Background Assistants" backed by a credit pool — treat those two specifically as reported-but-not-primary-confirmed until re-checked.

Best for

Find papers
Automate workflow

Also useful for

Talk through ideas

What it does well

Five frontier models (sonar-pro, GPT-5.4, Claude Opus 4.8, Gemini 2.5 Pro, Grok 4) bundled under one price, confirmed directly from Perplexity's own plan comparison.
Every answer still carries Perplexity's core strength — clickable source citations — regardless of which model answered.
Unlimited Pro Search and unlimited Labs remove the usage caps that limit the free/Pro tiers.
Single vendor relationship if you're already a Perplexity Pro user — no new account, billing, or data controller to onboard.

Where it falls short

The most expensive tool in this chapter ($200/mo vs $10–25/mo for the others) — a hard sell versus Abacus.ai or Genspark for the same "many models" pitch.
This session could not directly confirm the reported "Model Council" orchestration or "Background Assistants" automation features on Perplexity's own pricing page (it returned an access error) — verify these are still current before teaching them as fact.
No explicit "we don't train on your data" statement was found for Perplexity in this session's research — verify current policy directly before recommending it for sensitive material.
Overlaps heavily with the free-standing `perplexity` entry already taught in Frontier chatbots — only worth the separate purchase for Max-specific features.

Try it in 5 minutes

Compare answers across Max's bundled models on one question

Ask Perplexity Max the same research question, then check the model picker to see which of sonar-pro / GPT-5.4 / Claude Opus 4.8 / Gemini 2.5 Pro / Grok 4 answered, and switch models on the same thread.

Expected: A cited answer plus visible confirmation of which bundled model produced it — the "one price, several frontier models" idea made concrete inside a product students may already know.

Pricing

$200/month or $2,000/year (17% savings paying annually). Confirmed directly: 5 bundled models (sonar-pro, gpt-5.4, claude-opus-4.8, gemini-2.5-pro, grok-4), unlimited Pro Search, unlimited Labs, full Comet access, early access to new models, priority support. Reported but not confirmed: a Model Council orchestration feature and Background Assistants with a ~10,000-credit monthly pool.

⚠️ Watch out

At $200/mo it is the most expensive tool in this chapter by far — only worth it once you are already using Perplexity daily and specifically want its Background Assistants or multi-model Council, not as a first aggregator to try.

Code & automate

Claude Code Terminal agent that reads, writes, and runs your codebase

Claude Code

paid claude.ai/code →

by Anthropic (Claude team) · released Feb 2025

Terminal agent that reads, writes, and runs your codebase

Claude Code is Anthropic's agentic coding tool that runs in your terminal, VS Code, JetBrains, or browser and interacts with your actual codebase — not a sandboxed snippet. Released in February 2025 as a preview and reaching general availability in May 2025, it can read every file in a project, make coordinated multi-file edits, run tests, create git commits, open pull requests, and pipe output from other CLI tools. For biology students it means: paste a raw CSV, describe the analysis you need, and Claude Code writes the Python or R, runs it, and returns results — all from your terminal. It also supports scheduled routines (nightly batch jobs, recurring analysis runs) and can coordinate multiple parallel sub-agents on large tasks. The tool has grown rapidly; Anthropic reported a 5.5x revenue increase by July 2025 and non-programmers adopted it heavily during late 2025.

Best for

Write code
Automate workflow

Also useful for

Build an app
Draft docs

What it does well

Reads and edits multiple files in a single turn — ideal for refactors spanning several scripts or modules.
Runs your code directly and iterates on failures: if a seaborn plot crashes, it reads the traceback, fixes the bug, and re-runs.
Git-aware: stages changes, writes commit messages, creates branches and PRs without leaving the terminal.
Composable with Unix pipes — pipe log files, diff output, or CSV columns directly into a Claude Code command.
Supports MCP (Model Context Protocol) to pull in external data like Google Drive docs or Jira tickets inside the coding session.

Where it falls short

No free tier — a paid Claude subscription is mandatory to run it at all.
Fast mode costs ~6x the normal token rate; large repos or long sessions on Max plans can burn through credits unexpectedly.
Context window limits on very large codebases (>500k lines): Claude may not hold the full codebase in memory for a single turn and can lose earlier context in long sessions (a known bug, since patched but still occasional).
No native live web search — it cannot look up a PubMed abstract or fetch a protocol from a URL unless you explicitly paste the text.

Try it in 5 minutes

Generate a heatmap from CSV data

cd ~/lab-data && claude "Read measurement_data.csv and generate a heatmap with seaborn showing sample vs. condition. Save the figure as heatmap.png."

Expected: Claude writes a Python script, runs it, and saves heatmap.png in the same directory.

More prompts to try (2)

Write unit tests for an existing analysis script

claude "Write pytest tests for normalize_counts.py — cover edge cases like zero-count rows and missing columns."

Automate renaming a folder of raw sequencing files

claude "All .fastq.gz files in ./raw/ follow the pattern SAMPLE_LANE_R1.fastq.gz. Rename them to SAMPLE_R1_LANE.fastq.gz using a shell script, then run it and confirm no files were missed."

Pricing

Requires a Claude Pro ($17/mo annual or $20/mo monthly) subscription at minimum. Max plans at $100/mo (5x usage) and $200/mo (20x usage) exist for heavier workloads. "Fast mode" (priority inference) costs 6x more tokens per call — monitor usage carefully on Max plans.

Open the full Claude Code course 13 lessons · cheat sheet

⚠️ Watch out

Requires a paid Claude subscription ($17–20/month minimum); it writes files directly, so always start it in the correct project folder.

🏗 Building complex codebases A repeatable workflow for shipping features in a large codebase with an AI agent

🏗

Building complex codebases

free github.com/coleam00/ai-transformation-workshop →

by Methodology popularised by Cole Medin (Dynamous) and the agentic-coding community · released 2026

A repeatable workflow for shipping features in a large codebase with an AI agent

Building complex codebases is the practice of getting an AI coding agent to work reliably on large, real projects — not by buying a smarter model, but by giving it the right context. The core idea, popularised by Cole Medin, is that every codebase needs a second "AI layer" checked into source control alongside the code: a lean CLAUDE.md of project rules, reusable slash commands, and (when warranted) skills, subagents and MCP connections. You drive each task with a tight loop — Plan, Implement, Validate (the "PIV loop"): load just the right context, write a plan complete enough to run in a fresh window, implement, then validate against checks that have to pass. The discipline turns AI coding from unreliable one-offs into a repeatable engineering system that improves over time, because every mistake becomes a fix to the system, not just the bug. It works with Claude Code, Cursor, Codex or any agentic coding tool.

Best for

Write code
Automate workflow

Also useful for

Build an app
Data analytics

What it does well

Treats the AI layer (CLAUDE.md, commands, skills) as code: checked into git, reviewed, and improved like any other part of the repo.
The PIV loop (Plan → Implement → Validate) gives every task a fresh-context plan and a real check, so the agent cannot "talk its way to done".
Reusable slash commands turn repeated prompts into one-word workflows the whole team shares.
Errors become signals: each mistake is fixed in the system (a rule, a command, a check) so it does not happen twice.
Tool-agnostic — the same habits apply to Claude Code, Cursor, Codex or any agentic coding tool.

Where it falls short

Upfront cost: writing the AI layer and commands takes real time before it pays off — overkill for a tiny script.
A model can only reliably follow so many rules — a bloated CLAUDE.md hurts more than it helps, so it needs pruning.
Easy to overengineer: adding skills, subagents and MCP servers you do not need wastes context and tokens.
It is a practice, not a product — the discipline only works if you actually keep doing it.

Try it in 5 minutes

Write a lean CLAUDE.md for one real project

Open your project in Claude Code and ask: "Read this codebase and draft a CLAUDE.md with the build/test/lint commands, the directory map, and ~10 hard rules that would stop you making a mistake here. Keep it under 100 lines — only rules I would actually be wrong without."

Expected: A short, high-signal rules file that loads every session — so you stop re-explaining your project on each new task.

More prompts to try (2)

Run one task through Plan → Implement → Validate

Pick a small feature. First ask Claude to PLAN it (no edits): list the steps and the checks that must pass. Then in a fresh session, ask it to IMPLEMENT the plan. Finally ask it to VALIDATE: run lint, type-check and tests, and report what failed.

Commandify a repeated prompt

Take a prompt you retype often (e.g. "review this diff for security issues") and save it as a reusable command in your project's .claude/commands/ folder, then invoke it with a slash command.

Pricing

The workflow itself is free — it is a set of habits and version-controlled files. You only pay for the coding agent you run it on (e.g. Claude Code on a Claude Pro/Max plan, or a BYOK tool). Cole publishes the full reference setup — commands, rules templates and a demo app — as open source.

Open the full Building complex codebases course 6 lessons · recipes

⚠️ Watch out

This is a discipline you layer on top of a coding agent (Claude Code, Cursor, Codex), not a product you install. The payoff is real but it asks for upfront setup and steady habits — skip the discipline and you are back to one-off "vibe coding".

🍵 Gitea The self-hosted GitHub — own the server, own the code

🍵

Gitea

free open source about.gitea.com →

by Gitea Ltd. / the go-gitea open-source community · released 2016

The self-hosted GitHub — own the server, own the code

Gitea is a self-hosted alternative to GitHub: you install it on your own server (or a cheap VPS) and get the same core workflow — git repos, pull requests, issue tracking, code review — without a single commit leaving your own infrastructure. It ships as one lightweight Go binary or Docker image, so it runs comfortably on hardware that GitHub-scale tools would never fit on. Gitea Actions gives you GitHub-Actions-compatible CI/CD, so workflow files mostly copy over unchanged, meaning you can run AI coding agents, test suites, and deploy pipelines entirely on infrastructure you control. It also bundles a package registry (npm, PyPI, Docker images, and more) so a lab or small team can host code, packages, and CI in one place. For biology students, the practical case is control: unpublished thesis code, sensitive datasets, or a pre-print analysis pipeline stay on a server you or your institution owns, not a third-party cloud.

Best for

Write code
Open source / self-host

Also useful for

Build an app
Automate workflow

What it does well

Single Go binary — installs and upgrades in minutes, runs on a $5/month VPS
Gitea Actions is GitHub-Actions-compatible, so CI workflows port over with little to no rewriting
Built-in package registry (npm, PyPI, Cargo, Docker/Container, Maven, and 20+ more) alongside git hosting
Full data ownership — no vendor can see, train on, or restrict access to your repos
Free and open source (MIT) with no per-seat cost when self-hosted

Where it falls short

No built-in AI code-review or Copilot-style assistant — you bring your own agent or tooling
You (or your institution's IT) own uptime, backups, and security patching — no vendor safety net
Smaller ecosystem of third-party integrations and marketplace apps than GitHub
Governance split: in Oct 2022 the lead maintainers moved Gitea's trademark to a for-profit company, Gitea Ltd., which pushed part of the community to fork it into Forgejo (nonprofit-governed, under Codeberg e.V.) two months later — worth knowing before picking a side

Try it in 5 minutes

Push a script to a Gitea repo

Ask your instructor or IT for access to an existing self-hosted Gitea instance (or sign up free at gitea.com), create a new repository, then run `git remote add gitea <url>` and `git push gitea main` on a small analysis script.

Expected: Your code now lives on a git forge you or your institution controls, with full commit history and a web UI — no GitHub account involved.

Pricing

Self-hosted Gitea (the software) is fully free — MIT license, no seat limits, unlimited repos and users. Self-managed Enterprise (SAML SSO, audit logs, Kubernetes autoscaling runners, SLA support) is $9.50-19/user/month, where $9.50 needs a 1-year commitment, with a 30-day trial. Gitea Cloud, the managed hosting, has no permanent free tier — a 30-day trial, then a quoted plan, with discounts for startups and small teams. Signing up on the public gitea.com instance is still free.

Open the full Gitea course 5 lessons

⚠️ Watch out

Self-hosting means you are the sysadmin — backups, upgrades, and security patches are your job, not a vendor's.

🐙 GitHub Where the world's code lives and ships

🐙

GitHub

free tier github.com →

by GitHub (Microsoft) · released 2008

Where the world's code lives and ships

GitHub is the platform where most of the world's code is hosted, reviewed, and shipped: repositories, pull requests, Issues, and GitHub Actions for CI/CD, plus Codespaces for a cloud dev environment you can open from any browser. It's also fast becoming the place where AI agents do real work, not just where humans write code. The Copilot coding agent takes a GitHub Issue, works in an isolated sandbox, and opens a pull request on its own; GitHub Actions can run AI steps as part of a build; and GitHub's official MCP server lets any MCP-compatible agent (Claude included) read and write repos, Issues, and PRs directly. The catch is the "proprietary incumbent" trade-off: GitHub is free to start and has enormous network effects (open-source projects, collaborators, integrations all assume you're there), but it's owned by Microsoft, closed-source, and there's no self-hosting your way out short of the expensive Enterprise Server tier.

Best for

Write code
Automate workflow

Also useful for

Build an app
Templates

What it does well

Largest code-hosting network effect — most open-source projects, collaborators, and integrations already assume GitHub
GitHub Actions gives every repo free CI/CD out of the box, no separate service to wire up
Fastest-growing home for AI coding agents: Copilot coding agent opens PRs from an Issue, and the official MCP server lets any agent read/write repos
Codespaces spins up a full cloud dev environment from the browser — no local setup
Student Developer Pack makes GitHub Pro and dozens of partner tools free for verified students (its Copilot offer is paused for new sign-ups — verify before a cohort relies on it)

Where it falls short

Proprietary and Microsoft-owned — you don't control the server your code and history live on
No real self-hosting on Free/Team; the self-hosted option (GitHub Enterprise Server) is an expensive, separate product
Actions minutes are capped on Free/Team; heavier CI usage means either upgrading or paying per-minute overages
Workflows, Actions YAML, and agent integrations are written against GitHub's specific APIs — migrating away later is real work

Try it in 5 minutes

Put a lab analysis script under version control and open your first pull request

1) Create a free GitHub account and a new repository. 2) Push a small analysis script (e.g. an R or Python file). 3) Create a branch, make one small edit (fix a comment or add a docstring), and open a pull request against main. 4) Read the PR diff view — this is how every collaborator, human or AI agent, proposes a change on GitHub.

Expected: A public repo with one merged pull request, and a working sense of the Issue → branch → PR loop that Copilot's coding agent also uses.

Pricing

Free: unlimited public/private repos, unlimited collaborators, 2,000 Actions minutes/mo. Team: $4/user/mo for the first 12 months, adds org controls and 3,000 shared Actions minutes. Enterprise: from $21/user/mo, also an introductory rate for the first 12 months. Copilot is billed separately (free tier up to $100/user/mo for Copilot Max). GitHub Student Developer Pack still gives verified students GitHub Pro free, but its Copilot offer is paused for new sign-ups as of July 2026 — check education.github.com/pack before promising it to a cohort.

Open the full GitHub course 5 lessons

⚠️ Watch out

Proprietary and Microsoft-owned — your repos, Issues, and Actions workflows live on someone else's server, and there's real lock-in once your CI/CD is written in GitHub's YAML dialect.

🔁 Test-Driven Prompt Engineering Write your examples first, then edit the prompt until they pass

🔁

Test-Driven Prompt Engineering

free promptfoo.dev →

by Practice borrowed from software's Test-Driven Development; modern tooling includes the open-source promptfoo project · released 2026

Write your examples first, then edit the prompt until they pass

Test-Driven Prompt Engineering (TDPE) borrows a habit from software's Test-Driven Development and applies it to prompts: decide what "right" looks like before you write the prompt, not after you admire one good answer. You start with a handful of real `input → expected` pairs — your examples, including the edge cases and past failures — before touching the prompt itself. Then you run the red → green → refactor loop: find the example your draft prompt fails ("red"), make the smallest edit that fixes it ("green"), and once every example passes, tighten the prompt for clarity and cost without breaking any of them ("refactor"). The payoff is that you re-run the same example file whenever you edit the prompt, swap models, or a provider updates the model under you — catching a silent regression before your users do. The modern anchor for this is the open-source `promptfoo`, which keeps examples in a file and re-runs them on command, but a spreadsheet and a manual pass is a perfectly valid start. The ideas outlast any one tool.

Best for

Draft docs
Automate workflow

Also useful for

Talk through ideas
Write code

What it does well

Catches silent regressions — a prompt edit or model swap that breaks a case you already solved shows up immediately, not after a user complains.
Turns "the demo looked good" into a checkable spec: a small file of real input/output pairs anyone on the team can read and extend.
Makes model swaps safe — re-run the same examples against a new model or provider before you trust it in production.
Compounds over time: every failure you have ever seen becomes a permanent example, so the prompt gets harder to accidentally break.

Where it falls short

Needs you to write examples up front — skipped under deadline pressure, which is exactly when it is most needed.
Judging whether an output is actually "right" is its own skill, especially for open-ended text — see the Evals & testing chapter for grading (code checks, LLM-as-judge, human review).
A tiny example set still misses real-world inputs you never thought to write down — it lowers risk, it does not eliminate it.
Not a substitute for monitoring real usage — examples are a proxy for production, not a replacement for watching it.

Try it in 5 minutes

Write five examples before you touch the prompt

Pick a task you already use an AI assistant for (e.g. classifying an email, summarising an abstract). Before editing your prompt, write down 5 real `input → expected output` pairs — include at least one tricky edge case. Then run your current prompt against all five and mark each pass/fail.

Expected: A short list of input/expected pairs and a pass/fail column — your first spec, and probably at least one surprise failure.

More prompts to try (2)

Drive one fix from a red example

Take the example that failed above. Make the smallest possible change to your prompt (one added instruction, one example, one delimiter) that turns it green — then re-run all five examples to confirm you did not break the ones that were passing.

Refactor under green

Once every example passes, try shortening the prompt by a third for lower cost and clarity. Re-run the examples after the edit — if any turn red, you cut something load-bearing; put it back and try a smaller cut.

Pricing

The discipline itself is free — it is a habit and an example file. promptfoo is open source and free to run locally; its optional hosted/enterprise tier adds team features on top.

⚠️ Watch out

This is a discipline you layer on top of any prompt, not a product you install. It asks you to write examples before you tune — skip that and you are back to tweaking by vibes.

🚀 From demo to production agent What separates a real agent from a demo

🚀

From demo to production agent

free github.com/vercel/eve →

by Framework-agnostic practice; illustrated here with Vercel's eve (Vercel, Inc.) · released 2026

What separates a real agent from a demo

A demo agent is a prompt and a loop — type something in, watch it respond. A production agent is one you can trust to run unattended: on a schedule, over Slack, touching real data, while you are asleep. This chapter is that gap, and the pattern modern frameworks use to close it, illustrated with Vercel's open-source eve (Apache-2.0, public beta since June 2026). Eve's idea: an agent is a directory you can read, where file placement decides capability — instructions.md is the always-on system prompt, skills/*.md are markdown playbooks loaded on demand, tools/*.ts are the typed functions it can call, and subagents/, channels/, schedules/ and connections/ wire in delegation, entry points, triggers and external systems (Slack, OAuth, MCP). The smallest agent is two files — a model and its instructions. The parts a demo never needs and production always does then come built in rather than bolted on: durable execution (every conversation is a checkpointed workflow that survives a crash or deploy and resumes), a sandbox per agent, human approvals on risky actions, scored evals, a way in (Slack or HTTP, on a schedule), and a one-command deploy. The framework will change; the checklist will not.

Best for

Build an app
Automate workflow
Learning resources

Also useful for

Write code
Open source / self-host

What it does well

Convention over configuration: file placement IS capability — read instructions.md, skills/ and tools/ and you know exactly what the agent can do, no hidden graph or config object to trace.
Markdown-first lowers the barrier: instructions and skills are prose a non-programmer can read and edit; only the typed tools/*.ts need real code.
Production plumbing comes built in, not bolted on: durability, a sandbox, human approvals and scored evals are the things a demo never needs and production always does.
One command ships it — because an eve agent is an ordinary Vercel project, `vercel deploy` takes it from your laptop to production unchanged.
Framework-agnostic lesson: even if you never touch eve, "durability, evals, approvals, sandbox, a way in, deploy" is the checklist to run any agent framework through before trusting it.

Where it falls short

Still developer-oriented: instructions and skills are prose, but a real tool (tools/*.ts) is TypeScript — there is no code-free path to a fully custom agent.
Eve itself is young — public beta since June 2026, in a field where frameworks (and their APIs) are still moving fast; expect breaking changes.
The six production concerns catch known failure modes, not novel ones — durability and evals do not remove the judgement call of what is worth automating unattended.
The ideas outlast any one tool: eve is this chapter's worked example, not its subject — a framework favourite in 2026 may not be the one your lab uses in two years.

Try it in 5 minutes

Audit an agent for production-readiness

Here is a simple AI agent that answers questions from our lab wiki. Walk me through the demo→production checklist — durability, evals, approvals, sandbox, a way in, deploy — and tell me what is missing before I could trust it to run unattended.

Expected: A point-by-point gap analysis naming each production concern and what to add.

More prompts to try (2)

Read an agent as a folder

Look at the file tree of an eve agent (github.com/vercel/eve has examples). Find instructions.md, one file under skills/, and one file under tools/. For each, write one sentence: what would break if this file were deleted?

Map your own idea onto the six production concerns

Pick an AI agent idea from your own work — even a simple one, like a bot that answers questions from a lab wiki. For durability, evals, approvals, a sandbox, a way in, and deploy: write one sentence on whether it applies and what you would need to add.

Pricing

The eve framework is free and Apache-2.0 licensed — clone it, self-host it, or run it anywhere Node.js runs (`npx eve@latest init`). The one-command `vercel deploy` path runs on your existing Vercel plan (a free Hobby tier exists); either way you separately pay for whichever model API the agent calls.

⚠️ Watch out

Framework-agnostic checklist illustrated with one framework (Vercel's eve, itself a young public beta) — the six concerns matter whichever agent framework you actually use.

🤖 AI for robotics & edge devices The same AI assistants, applied to firmware, robots and constrained hardware

🤖

AI for robotics & edge devices

free developer.nvidia.com/embedded/jetson-modules →

by A synthesis of the robotics + edge-AI ecosystem (ROS 2, TensorFlow Lite Micro, ONNX/TensorRT, NVIDIA Jetson, LeRobot) — vendor-neutral · released 2026

The same AI assistants, applied to firmware, robots and constrained hardware

A domain chapter for robotics, mechatronics and embedded engineers who know their field but are new to AI: how everyday AI assistants speed up real work — writing firmware and ROS 2 code, getting a model small enough to run on the device, and adding an LLM/VLM reasoning layer to a robot — and exactly where they stop helping. It applies the coding-assistant and model-choice skills from the earlier chapters to microcontrollers, edge accelerators and single-board computers, and is honest about the immature parts (LLM-as-robot-brain in production, and raw HDL generation).

Best for

Write code
Automate workflow

Also useful for

Data analytics
Build an app

What it does well

Meets a robotics engineer where they are: C/C++/Rust, ROS 2, microcontrollers and edge boards — not generic web examples.
Covers the part that is genuinely a workflow win: the conversion + quantization plumbing to get a trained model onto constrained hardware.
Separates the real (sim-to-real, perception, VLA research) from the hype (an LLM as the production robot brain) so you do not chase demos.
Honest about the FPGA limit — points you at HLS you verify rather than letting a model promise you Verilog.
Builds on the existing coding-assistant and model-choice chapters instead of re-teaching them.

Where it falls short

It assumes you already have the domain expertise — it speeds up an engineer, it does not turn a non-engineer into one.
On-device ML (TinyML, quantization) is a distinct skill the assistant helps with but does not replace — you still measure before and after.
The robot foundation-model frontier (VLAs, large-scale sim-to-real) is mostly the territory of well-resourced labs, not a solo workflow yet.
Safety-critical and certified paths (IEC 60601, ISO 13485, MISRA) still need human verification and process — AI output is a draft, not a sign-off.

Try it in 5 minutes

Get a coding assistant to read your firmware repo

Open your firmware project in a coding assistant and ask: "Read this embedded codebase. Summarise the architecture (HAL, drivers, control loop, RTOS tasks), and list the 8 hard rules you must never break here — ISR constraints, blocking calls, memory, timing — so you do not make a dangerous change."

Expected: A grounded map of your firmware plus a short rules list you can drop into the project so every later request respects the constraints.

More prompts to try (2)

Quantize a model to fit the device

Ask: "I have a trained classifier and an MCU with 256 KB RAM / 1 MB flash. Walk me through INT8 post-training quantization, export it to TFLite-Micro as a C array, and estimate the tensor-arena size. Flag any ops the Micro runtime will not support."

Pressure-test the FPGA limit yourself

Ask: "Draft a Vitis HLS C++ kernel for a FIR filter on my sensor stream, with a testbench — and be explicit about what you can NOT guarantee about timing closure and resource usage." Notice how much it hedges: that is the honest boundary.

Pricing

The practices are free — they are habits plus open tooling (TFLite-Micro, ONNX, ROS 2, vendor SDKs). You pay only for the coding assistant you run them on and for whatever hardware you target.

Open the full AI for robotics & edge devices course 4 lessons · recipes

⚠️ Watch out

AI assistants augment robotics work — they do not replace the control theory, the standards, or the bench. They are strong on boilerplate, drivers, refactors and the model-to-device plumbing, weak on anything safety-critical you have not verified, and not trustworthy at all for production FPGA HDL.

🧩 Skills, tools & extensions The extensibility layer every agent shares — skills, tools, subagents & hooks

🧩

Skills, tools & extensions

free agentskills.io →

by Agent Skills is an open standard originated at Anthropic; MCP (Model Context Protocol) is an open standard from Anthropic + community · released 2025

The extensibility layer every agent shares — skills, tools, subagents & hooks

Every AI agent is a small core plus an extensibility layer, and the same five pieces show up everywhere. A skill is a folder with a SKILL.md file (a name, a description, and instructions) that the agent loads only when a task matches — so specialised know-how does not bloat every session. A slash command is a saved prompt you invoke by name. A subagent is a helper that runs in its own isolated context. An MCP tool connects the agent to an external system (GitHub, a database, your calendar) over the open Model Context Protocol. A hook is a script that fires automatically on an event. The skill format is an open standard — Agent Skills — that began at Anthropic and is now read by dozens of tools (Claude Code, Cursor, Codex, OpenCode, Goose, Copilot, Gemini CLI and more), so one SKILL.md works across all of them. This chapter is the tool-agnostic hub: the mental model, a decision guide for which piece to reach for, where to find ready-made ones, and how to write your own.

Best for

Automate workflow
Learning resources

Also useful for

Write code
Build an app

What it does well

One mental model for any agent: skills, commands, subagents, MCP tools and hooks recur across Claude Code, Hermes, Cursor and the rest.
Portable by design: the Agent Skills format (a folder + SKILL.md) is an open standard, so a skill you write is read by dozens of tools unchanged.
Progressive disclosure keeps it cheap: an agent loads only a skill's name and description until a task needs it, then reads the full instructions — so many skills cost little context.
A clear decision guide: reusable know-how → skill; a shortcut → command; isolate noisy work → subagent; reach an external system → MCP tool; auto-fire on an event → hook.
Rich, browsable catalogs already exist — Anthropic's skills repo, the agentskills.io showcase, and the official MCP registry — so you rarely start from zero.

Where it falls short

It is a concept layer, not a product: the actual menus and file paths differ per tool, so you still read your agent's own docs to wire it up.
Easy to over-add: every skill, subagent and MCP server costs context and tokens, so restraint beats collecting — add only on a real need.
Quality varies in community catalogs — a downloaded skill or MCP server is someone else's instructions/code, so review it before trusting it with your data.
MCP tools and hooks can take actions, so loose permissions are a real risk — grant the narrowest access that works.

Try it in 5 minutes

Write a one-file skill and watch the agent pick it up

Make a folder with a single SKILL.md: a `name`, a one-line `description` of when to use it, and 5–10 lines of instructions for a task you repeat (e.g. "format a citation in our house style"). Drop it where your agent reads skills, then ask for that task in plain language.

Expected: The agent recognises the task from your description, loads the skill, and follows your instructions — without you pasting them in.

More prompts to try (2)

Browse a catalog and install one skill

Open agentskills.io or github.com/anthropics/skills, find a skill close to something you do, read its SKILL.md, and add it to your agent. Notice how the name + description tell the agent when to use it.

Connect one MCP tool

Browse the official MCP registry (registry.modelcontextprotocol.io), pick a low-risk server (e.g. a filesystem or fetch server), and connect it to your agent so it can read real data — then ask it a question only that data can answer.

Pricing

The concepts and formats are free and open. Skills and MCP servers are plain files you write or download; you only pay for whatever agent and model you run them on (e.g. Claude Code on a Claude plan, or a self-hosted assistant like Hermes on your own model).

Open the full Skills, tools & extensions course 4 lessons · recipes

⚠️ Watch out

This is a concept layer, not a product you install — the actual buttons live inside your agent (Claude Code, Hermes, Cursor…). The payoff is portability: learn it once and it carries between tools, but each tool has its own wiring details.

Claude Cowork Shared skills and plugins for Claude Code teams

Claude Cowork

paid claude.ai →

by Anthropic + community plugin authors · released 2025

Shared skills and plugins for Claude Code teams

Claude Cowork is the plugin and skill-sharing layer built on top of Claude Code. Where Claude Code is a personal coding agent, Cowork makes it a team tool: organisations author shared "skills" (reusable slash-command behaviours), wire Claude into their specific apps and APIs via MCP connectors, and distribute role-matched configurations to every developer on the team. A skill might be a deployment runbook, a code-review protocol, or a specialised data-pipeline agent — once published, any team member invokes it with a single slash command. Cowork also provides structured coauthoring workflows (documentation, specs, proposals) and a plugin registry for connecting Claude to external services (Slack, GitHub, Linear, Jira, databases). For a biology lab it means: one researcher authors a "literature review" skill, and every student on the project inherits it instantly. The plugin system is open for community contributions.

Best for

Learning resources
Templates

Also useful for

Write code
Draft docs
Automate workflow

What it does well

Shared skills: author a slash-command once and every team member gets it — no copy-pasting prompts.
MCP connectors bridge Claude to external services (GitHub, Slack, databases, REST APIs) without writing glue code.
Role-matched setup: a "biology lab" plugin installs the right tools, rules, and skill presets in one command.
Coauthoring workflows for docs, specs, and proposals with structured review checkpoints.
Plugin registry: community-contributed connectors cover dozens of services out of the box.

Where it falls short

Plugin ecosystem is early (2025) — coverage and stability vary across connectors.
Requires Claude Code familiarity; not a beginner tool.
Plugin authoring (writing SKILL.md files, MCP server configs) has a learning curve.

Try it in 5 minutes

Install the setup-cowork skill

In Claude Code, type: /setup-cowork
Follow the prompts to install role-matched plugins for your workflow.

Expected: Claude installs matching plugins, connects your tools, and suggests a first skill to try — tailored to your role.

More prompts to try (1)

Create a team skill

Ask Claude Code: "Help me create a skill that runs our lab's standard literature review workflow — PubMed search + summary + citation list."

Pricing

Included with any Claude Code subscription (Pro $17-20/mo, Max $100-200/mo). Plugins themselves are free and open-source; you pay only for the Claude tokens consumed.

⚠️ Watch out

Requires a Claude Code subscription; the plugin ecosystem is still early — expect rough edges and evolving APIs.

Hermes Self-hosted AI agent that remembers you and writes its own skills

Hermes

free open source hermes-agent.nousresearch.com →

by Nous Research · released 2026

Self-hosted AI agent that remembers you and writes its own skills

Hermes Agent is a self-hosted, open-source autonomous AI agent from Nous Research — "the agent that grows with you." Unlike a chat window you start fresh each time, Hermes keeps a persistent memory of your preferences, projects, and environment across every session, so you stop re-explaining context. When it solves a hard problem it writes a reusable "skill" document (compatible with the agentskills.io standard) so it never forgets how to do it again. You reach it through the messaging apps you already use — Telegram, Discord, Slack, WhatsApp, Signal — or the CLI, all from one gateway process, and a built-in cron scheduler lets it run unattended tasks (daily digests, nightly backups, weekly audits) described in plain language. It runs on local models for full privacy or cloud models for more power, and deploys six ways (local, Docker, SSH, Singularity, and the serverless Modal/Daytona backends that hibernate when idle). For a researcher it is the tool when you want a private, always-on assistant that accumulates context about your work instead of forgetting it. (Note: distinct from the Hermes large language models, also by Nous Research — this is the agent that runs on top of a model.)

Best for

Automate workflow
Talk through ideas

Also useful for

Write code
Open source / self-host

What it does well

Persistent memory — remembers your preferences, projects, and past work across every session, so you never re-explain context.
Writes its own reusable skills — turns a solved problem into a saved, shareable skill (agentskills.io standard) it can reuse later.
Lives in your messaging apps — one gateway connects Telegram, Discord, Slack, WhatsApp, Signal, and the CLI.
Autonomous scheduling — a built-in cron runs unattended tasks (daily digests, backups, audits) described in plain language.
Model-agnostic and private — runs on local models so nothing leaves your machine, or cloud models when you need more power.

Where it falls short

You self-host and maintain it — install, connect a model, and keep a process running; not a click-and-go web app.
Quality depends entirely on the model you connect — a small local model limits what the agent can actually do.
Young, fast-moving project (v0.x) — expect rough edges and occasional breaking changes.
An always-on agent with memory and tool access needs supervision — be deliberate about what it may do and remember unattended.

Try it in 5 minutes

Install Hermes and give it a memory

1. Install: curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
2. Connect it to Telegram (or just use the CLI).
3. Tell it: "Remember that I work on zebrafish development and prefer answers with citations."
4. End the session, start a new one, and ask a related question — watch it apply what it remembered.

Expected: On the second session it recalls your context without being re-told — the core "grows with you" behaviour.

More prompts to try (3)

Schedule an unattended daily digest

Ask Hermes: "Every weekday at 8am, summarise any new arXiv preprints on zebrafish neurodevelopment and send the summary to me on Telegram."

Give a project its own remembering agent

In a Slack channel for your thesis, tell Hermes: "This channel is for my RNA-seq thesis — remember the decisions we make here and post me a status digest every Monday morning."

Teach it a skill it can reuse

Walk Hermes through cleaning one messy CSV (drop empty rows, fix headers, export tidy). Then say: "Save that as a skill called tidy-csv." Next time just say "run tidy-csv on this file."

Pricing

Free and open-source — you only pay for the model it runs on. Point it at a local model (Ollama / LM Studio) for $0 in tokens and full privacy, or connect a cloud model (Claude, GPT) and pay that provider per token. The serverless Modal/Daytona backends cost almost nothing while the agent is idle.

Open the full Hermes course 9 lessons · cheat sheet · recipes · downloads

⚠️ Watch out

You self-host it — expect a real setup (install script, a model to connect, and a server or laptop to keep it running). An always-on agent with memory and tool access also needs care: review what you let it do unattended. It is a young, fast-moving project (v0.x).

Antigravity Agent-first dev platform — orchestrate many AI agents at once

Antigravity

free tier antigravity.google →

by Google · released Nov 2025 (2.0 at I/O, May 2026)

Agent-first dev platform — orchestrate many AI agents at once

Antigravity is Google's agent-first development platform: instead of one assistant in a chat box, it gives AI agents their own space to work. A "Manager Surface" lets you spawn, orchestrate, and watch multiple agents running in parallel across different workspaces, spin up subagents to split a job, and schedule tasks to run in the background on a cron. An "Editor View" is a full AI IDE with tab-completion and inline commands. It ships as a desktop app (Mac/Windows/Linux), a CLI, an SDK, and a managed agent service, and it is model-agnostic — Gemini 3 Pro, Anthropic's Claude Sonnet 4.5, and OpenAI's GPT-OSS all run inside it, with live voice transcription via Gemini Audio. For a researcher it suits bigger, multi-step jobs you want to hand off — refactor a pipeline, build a tool, run an overnight task — and check on later, rather than one quick question.

Best for

Write code
Automate workflow

Also useful for

Build an app

What it does well

Orchestrate many agents at once — the Manager Surface spawns, schedules, and observes parallel agents across workspaces.
Background + scheduled work — hand off a job (or a cron task) and let agents run while you do something else.
Model-agnostic — switch between Gemini 3 Pro, Claude Sonnet 4.5, and GPT-OSS inside one platform.
Real toolchain — desktop app, CLI, SDK, and a managed agent service, on Mac/Windows/Linux.
Full AI IDE — the Editor View has tab-completion and inline commands alongside the agents.

Where it falls short

Google account + sign-in required; you are on Google's platform, not self-hosted.
Free tier is capped (~20 agent requests/day) — a real prototyping session exhausts it quickly.
Young and fast-moving — 2.0 was a full rebuild shipped May 2026; expect change.
High-volume use is pricey (Ultra up to $250/mo); agents act on your code and need supervision.

Try it in 5 minutes

Hand a whole task to an agent

In Antigravity, open a new agent and ask: "Write a Python script that fetches a UniProt entry by accession, prints the protein name, length, and organism, then add a test and run it — fix any errors until the test passes."

Expected: The agent plans, writes the script and test, runs it, and iterates on failures — you watch it work rather than copy-pasting code.

More prompts to try (2)

Run two agents in parallel

Spawn one agent to "refactor this analysis script into modules with type hints" and a second to "write pytest tests for the normalisation step" — watch both work at once in the Manager Surface.

Schedule an overnight job

Schedule a task: "Every night, pull the latest commits, run the full test suite, and post a summary of any failures to me in the morning."

Pricing

Free tier: $0, ~20 agent requests/day on Gemini Flash. AI Pro $20/mo (~100/day, better models). AI Ultra $100–$250/mo for high-volume use. Pay-as-you-go credit top-ups at $0.01/credit when quota runs out.

Open the full Antigravity course 7 lessons · cheat sheet

⚠️ Watch out

Needs a Google account and runs Google's platform; the free tier is capped (~20 agent requests/day) and heavy use gets expensive. Agents act on your code, so review what they do.

openclaw Run AI agents on your own devices via WhatsApp or Telegram

openclaw

free tier open source openclaw.ai →

by Peter Steinberger (@steipete), now maintained by openclaw foundation · released Nov 2025 (as Clawdbot)

Run AI agents on your own devices via WhatsApp or Telegram

openclaw (originally Clawdbot, Nov 2025; renamed Moltbot Jan 27 2026; then openclaw) is an open-source personal AI agent created by Peter Steinberger (@steipete), who announced on February 14, 2026, that he was joining OpenAI and the project would move to a foundation while remaining open-source. The project reached 247,000 GitHub stars by that date — a remarkable growth trajectory in under four months. The concept is simple: run an AI agent on your own machine and talk to it through messaging apps you already use — WhatsApp, Telegram, Signal, Discord, Slack, Google Chat, iMessage, or 15+ others. The agent can manage email, read and write calendar events, browse the web autonomously, control smart home devices, run shell scripts, and maintain a persistent memory that personalizes across sessions. It supports Claude, GPT, and local models via Ollama, so you can choose which AI model handles your tasks. The Go plan on opencode.ai is unrelated — openclaw itself is MIT-licensed and free.

Best for

Automate workflow
Talk through ideas

Also useful for

Draft docs

What it does well

Runs on your own hardware — your data never passes through a third-party cloud (beyond your chosen LLM API).
Talk to your AI through apps you already have: Telegram, WhatsApp, or Discord — no new app to learn.
Persistent memory: remembers your preferences, regular tasks, and lab schedule across conversations.
50+ integrations: Gmail, Google Calendar, GitHub, Spotify, Obsidian — connects real workflows, not just demos.
Cron-style proactive tasks: schedule it to send a daily digest of emails or a weekly summary of open GitHub issues.

Where it falls short

Setup is technical — requires Node 24+, a working API key, and running a local server process. Not beginner-friendly.
Voice mode is macOS/iOS-only currently (wake-word detection); Android has continuous voice only.
The project is in governance transition after Steinberger joined OpenAI — update frequency and leadership may shift.
No visual output by default — answers appear as text in your messaging app, so complex tables or charts require openclaw OS on top.

Try it in 5 minutes

Ask your personal agent to summarize today's emails

Send via Telegram to your openclaw bot: "Summarize all unread emails from today and flag anything urgent."

Expected: openclaw reads your inbox and replies with a bulleted summary directly in Telegram.

More prompts to try (1)

Schedule a weekly lab digest

Send: "Every Monday at 8am, summarize my open tasks in GitHub, any calendar events this week, and the 3 newest papers by [supervisor name] on PubMed."

Pricing

MIT-licensed and self-hosted — free to run, but you need your own LLM API key (Claude or OpenAI). Estimated cost: $5–15/month in API calls for moderate personal use. A managed cloud version was rumored at ~$0.66/day (~$20/month) — check openclaw.ai for current terms as the project is in active transition.

Open the full openclaw course 4 lessons · cheat sheet

⚠️ Watch out

Setup requires Node 24+, an LLM API key, and a local host machine — expect a 30–60 minute first-time setup, not a point-and-click tool.

opencode Open-source terminal agent — your key, any model, full control

opencode

API key required open source opencode.ai →

by Anomaly (open-source; Dax Raad and team, originally SST) · released 2025

Open-source terminal agent — your key, any model, full control

opencode is an open-source AI coding agent built by Anomaly (formerly known as the SST team, led by Dax Raad) that runs in your terminal, as a desktop app, or as an IDE extension. It reached 160,000+ GitHub stars and 7.5 million monthly developers as of May 2026, placing it among the most-adopted open-source coding agents. The core value proposition: bring your own API key from any of 75+ supported providers (Anthropic, OpenAI, Google, Mistral, or local models via Ollama), or connect your existing GitHub Copilot or ChatGPT subscription. You pay your provider directly — there is no intermediate markup. Two built-in agent modes switch with Tab: "build" executes code changes, "plan" explores and reads without modifying files. A $10/month "Go" plan provides generous access to curated open-source models (DeepSeek, Qwen, Kimi) without managing your own keys — useful as a cost-controlled entry point.

Best for

Write code

Also useful for

Automate workflow

What it does well

Model-agnostic: switch between Claude, GPT, Gemini, DeepSeek, or a local Ollama model in a single config line — no re-install.
Privacy-first design: no code or context stored server-side (only your provider's API receives data).
Multi-session parallelism: run two separate agent sessions on the same project simultaneously for different subtasks.
LSP-enabled: loads language server protocol automatically for better code understanding and navigation.
Composable with GitHub Copilot or ChatGPT Plus subscriptions — you can reuse a subscription you already pay for.

Where it falls short

BYOK means costs land directly on your API bill — a long refactoring session with Claude 3.7 can cost $5–20 if you're not careful.
Terminal-first UX: no visual diff or inline editing without installing the IDE extension — steeper learning curve than Claude Desktop.
No built-in live web search or citation lookup — purely a coding and file-editing tool.
Desktop app is still beta (macOS/Windows/Linux) and less mature than the terminal CLI.

Try it in 5 minutes

Debug an R script in the terminal

opencode "My R script anova_analysis.R throws an error on line 34. Read the file, find the bug, and fix it."

Expected: opencode reads the file, explains the error, and proposes a fix — or applies it automatically.

More prompts to try (1)

Generate a Python data-cleaning pipeline

opencode "Read all CSV files in ./data/ and write a script that merges them, removes duplicates by sample_id, fills missing numeric values with column medians, and saves the result as cleaned_data.csv"

Pricing

Core tool is free and MIT-licensed — you only pay your model provider (e.g., Anthropic API rates: ~$3/million input tokens for Claude 3.7). The "Go" plan costs $10/month (first month $5) and provides rate-limited access to open-source models like DeepSeek V4 and Qwen3.5 Plus without needing your own API key.

Open the full opencode course 6 lessons · cheat sheet

⚠️ Watch out

BYOK means token costs hit your Anthropic or OpenAI bill directly — monitor usage or cap spending to avoid surprises.

🎙️ opencode + voice Talk to your terminal coding agent — local speech in, local speech out

🎙️

opencode + voice

free open source crates.io/crates/opencode-voice →

by Community / local-first stack — the opencode-voice Rust CLI (independent), the Olbrasoft/opencode-voice-plugin, and any local TTS; demoed end-to-end by the YouTube channel "No place like localhost". · released 2026

Talk to your terminal coding agent — local speech in, local speech out

A hands-free way to drive opencode, the open-source terminal coding agent, entirely by voice — with speech recognition and voice output both running on hardware you control, not a cloud API. You start opencode with a local port open, then run a small Rust CLI (opencode-voice) that listens on your microphone, transcribes with Whisper on your own CPU, and forwards the text to opencode. For opencode to answer out loud, a community plugin adds a 'speak' tool the agent can call, routed to a text-to-speech engine — anything from CPU-friendly Piper to a GPU voice-cloning model. None of this is official SST tooling: it's independent open-source projects wired together via opencode's HTTP server and plugin system. For life-science students the appeal is privacy-by-architecture — review or fix code by voice near sensitive data without your spoken words hitting a third-party server — and a genuinely hands-free workflow when your hands are on a pipette, not a keyboard.

Best for

Talk through ideas
Write code

Also useful for

Automate workflow

What it does well

Speech recognition and speech output can both run on hardware you control — no audio leaves your network
Push-to-talk listening is one `cargo install` and a flag; Whisper runs CPU-only (~600MB model), no GPU needed
opencode's plugin system means the 'speak' tool is just another tool the agent can call — no core fork
Fully open-source throughout, and the TTS engine is swappable (CPU Piper up to GPU voice-cloning)
A real hands-free workflow for reviewing or fixing code away from the keyboard

Where it falls short

Not an official opencode feature — three independently-maintained community projects, any of which could go stale
Voice cloning needs a real GPU (~10GB VRAM), infeasible on a typical laptop; use CPU Piper for a laptop-friendly voice
Setup is fiddly: native build deps, a plugin symlink, and possibly your own TTS wrapper — not a five-minute install
"Local" describes the voice layer only — the LLM opencode calls may still be a cloud API unless you point it at a local model

Try it in 5 minutes

Ask opencode a question by voice

Start `opencode --port 4096`, run `opencode-voice --port 4096 --no-global` in a second terminal, hold space and ask: "Does the README mention a license?"

Expected: opencode-voice transcribes your question locally with Whisper and opencode answers in the terminal as if you had typed it.

Pricing

Every piece is free and open-source (opencode-voice is Apache-2.0; the plugin and opencode itself MIT). You supply the compute: a CPU is enough for the listening half; a GPU (~10GB VRAM) is only needed for voice cloning.

Open the full opencode + voice course 3 lessons

⚠️ Watch out

Not an official opencode feature — three separately-maintained community projects glued together; the voice-cloning route needs a real GPU (~10GB VRAM), though the listening half is CPU-only.

jcode Open-source terminal harness — any model, even local

jcode

API key required open source jcode.sh →

by Solo Systems (open source; creator Jehuang / GitHub @1jehuang) · released Jan 2026

Open-source terminal harness — any model, even local

jcode is an open-source terminal coding agent — a Rust "harness" built by Solo Systems (MIT-licensed, 8,600+ GitHub stars within months of its January 2026 launch) whose whole design is to squeeze the most useful work out of whatever model you point it at. The model-access story is what sets it apart: bring your own API key, reuse a Claude or OpenAI subscription you already pay for via OAuth, or run fully local models through Ollama or LM Studio — so your code and your costs can stay on your own machine. Beyond a normal chat agent it adds persistent memory (it keeps learning your project across sessions), side panels, background tasks, and swarm coordination — running several agents in parallel to scale test-time compute on a hard task. It speaks MCP, ships its own open "jcode bench", and installs with a single command on macOS, Linux, or Windows.

Best for

Write code

Also useful for

Automate workflow

What it does well

Model-flexible by design: bring your own key, OAuth into an existing Claude/OpenAI subscription, or run local models via Ollama or LM Studio — no markup.
Persistent memory carries context across sessions, so the agent keeps learning your project instead of starting cold each time.
Swarm coordination runs several agents in parallel to scale test-time compute on a hard task.
Side panels and background tasks keep long-running work non-blocking while you stay in the terminal.
Fully open source (MIT, Rust) with MCP support and its own open "jcode bench" — inspect it, self-host it, or fork it.

Where it falls short

Brand-new (Jan 2026) and moving fast — smaller ecosystem, fewer tutorials, and more churn than mature tools like Aider or Claude Code.
Terminal-only (TUI): comfortable if you live in the shell, less so if you want a graphical editor with inline diffs.
BYOK means model costs hit your own API bill; the optional hosted plans pre-pay usage but don't remove the need to watch spend.

Try it in 5 minutes

Fix a Python analysis script with a local model

jcode "Read cleanup.py, find why it drops rows with missing gene IDs, and fix it so those rows are kept and flagged instead."

Expected: jcode reads the file, explains the bug, and applies a fix — running on your own key or a local Ollama / LM Studio model.

Pricing

The jcode software is free and MIT-licensed with no feature gates — bring your own API keys, reuse an existing Claude or OpenAI subscription via OAuth, or run local models through Ollama or LM Studio at zero token cost. Optional hosted-token plans (Plus $10, Pro $20, Max $100, Ultra $200, Flagship $1000/mo) only pre-pay hosted model usage; they don't unlock features.

⚠️ Watch out

Very new (public since January 2026): fast-moving and rough in places, and with BYOK the token costs still land on your own model bill.

Codex OpenAI's terminal coding agent — reads, edits, runs your repo

Codex

paid developers.openai.com/codex →

by OpenAI · released 2025

OpenAI's terminal coding agent — reads, edits, runs your repo

Codex is OpenAI's agentic coding tool — a terminal agent (plus a VS Code extension and managed cloud tasks) that reads your repository, edits files, runs commands, and iterates on test failures until they pass. The CLI is open source (Rust) and runs locally in your own checkout, so your code stays on your machine; the models behind it are OpenAI's frontier models, billed through a paid ChatGPT plan or API credits. For a biology student it shines at refactoring a messy analysis script, adding a feature to a lab tool, or debugging a failing notebook end-to-end.

Best for

Write code
Automate workflow

Also useful for

Data analytics
Build an app

What it does well

Purpose-built for code — reads the repo, edits across files, runs the tests, and fixes what breaks.
Runs locally in your own checkout (the CLI is open-source Rust); your code never leaves your machine.
Three ways to run it: native terminal, VS Code extension, or fully managed cloud tasks.

Where it falls short

OpenAI models only — you cannot bring your own model or point it at a local one.
Needs a paid ChatGPT plan or API credits; no genuinely free tier.
Coding-focused — weak at open-web browsing or general chat versus a generalist agent.

Try it in 5 minutes

Refactor a lab analysis script

codex "Split my single-file RNA-seq script into modules, add type hints, and write pytest tests for the normalization step. Run the tests and fix any failures."

Expected: A refactored, tested repo — Codex edits files, runs the tests, and iterates until green.

Pricing

Included with ChatGPT Plus ($20/mo) and Pro plans, or pay per token via the OpenAI API. The CLI itself is free and open source; you pay for model usage.

Open the full Codex course 12 lessons · cheat sheet

⚠️ Watch out

Needs a paid ChatGPT plan or API credits, and runs OpenAI models only — no bring-your-own-model and no local-model option.

Claude Agent SDK Build your own agents on the engine behind Claude Code

Claude Agent SDK

API key required code.claude.com/docs/en/agent-sdk/overview →

by Anthropic · released 2025

Build your own agents on the engine behind Claude Code

The Claude Agent SDK is the engine behind Claude Code, exposed as a library so you can build your own agents on the same agent loop, built-in tools, and context management. Available in Python and TypeScript, it lets you ship a custom autonomous agent inside your own app — wiring in your own tools and logic while inheriting the planning and memory handling that make Claude Code work. The SDK is free; you pay Claude API tokens for what the agent does. For a lab it is how you would build, say, a protocol-checker agent that reads experiments and flags missing controls.

Best for

Build an app
Automate workflow

Also useful for

Write code

What it does well

Inherits Claude Code's agent loop, built-in tools, and context management — you don't rebuild the hard parts.
Python and TypeScript libraries that drop into software you ship.
Full tool-use / function-calling so the agent can act on your own systems.

Where it falls short

Requires coding — it's a library, not a point-and-click tool.
You pay Claude API tokens for every run, on top of building the agent.
More setup than Claude Code if all you need is a coding assistant.

Try it in 5 minutes

A protocol-checker agent

Build a small Python agent on the Agent SDK that reads an experimental protocol and flags missing controls, ambiguous steps, and unsafe reagents.

Expected: A reusable agent on the same engine as Claude Code, with your logic on top.

Pricing

The SDK is free and open source; usage is billed as Claude API tokens (Haiku/Sonnet/Opus tiers). Free trial credits available.

⚠️ Watch out

It's a developer library, not a ready-to-use app — you write the agent in Python or TypeScript and pay Claude API tokens for usage.

Claude API Call Claude models from your own code, billed per token

Claude API

API key required www.anthropic.com/api →

by Anthropic · released 2023

Call Claude models from your own code, billed per token

The Claude API is Anthropic's developer API: raw programmatic access to the Claude models from your own code. You call client.messages.create() via the anthropic SDK (Python or TypeScript) or the REST endpoint, choosing the model, shaping every prompt, and paying per token across the Haiku, Sonnet, and Opus tiers. It gives you the most control of any Claude product — and the least scaffolding: there is no agent loop or memory unless you build it. For a biology student it is the tool for batch jobs, such as classifying 500 PubMed abstracts by method into structured JSON from your own script.

Best for

Build an app
Automate workflow

Also useful for

Data analytics

What it does well

Full control — pick the model and shape every message; raw access with no wrapper.
Clean Python/TypeScript SDK plus a REST endpoint; embed Claude in anything you ship.
Ideal for batch and pipeline work — classify, extract, or summarize at scale into structured JSON.

Where it falls short

No built-in agent loop or memory — you build the app logic yourself.
Requires coding; not usable without writing a program.
Per-token billing adds up on large batches if you don't watch usage.

Try it in 5 minutes

Batch-classify 500 abstracts

Write a Python loop that calls messages.create() over 500 PubMed abstracts and tags each by method (RNA-seq, microscopy, …) as structured JSON.

Expected: A CSV of structured labels from your own script — full control over model and prompt, billed per token.

Pricing

Pay per token; pricing varies by model tier (Haiku cheapest, Opus most capable). Free trial credits to start; the anthropic SDK is free.

⚠️ Watch out

Raw programmatic access — you build the app logic and the agent loop yourself; there's no built-in planning or memory.

Ollama Run open models locally from the CLI + a local API

Ollama

free open source ollama.com →

by Ollama · released 2023

Run open models locally from the CLI + a local API

Ollama runs open large language models locally from the command line and exposes them through an OpenAI-compatible API on localhost:11434. Free and open source, it handles model download, quantization, and GPU acceleration for you, and runs as a long-lived service — a private endpoint your scripts, apps, or whole lab can call without sending data to the cloud or paying per token. It carries a curated model registry plus any GGUF model. For a biology student it is the way to give analysis scripts a local LLM: point your Python at localhost and have it summarize results or draft figure captions, fully offline.

Best for

Automate workflow
Data analytics

Also useful for

Talk through ideas

What it does well

OpenAI-compatible local API — swap one base URL and keep your existing code.
Runs as an always-on private service your scripts or team can hit; data never leaves your hardware.
Open source, with a big model selection (curated registry plus any GGUF).

Where it falls short

CLI/API-first — no graphical chat window of its own.
Local only; you can't call cloud models through it.
Quality and speed depend on your hardware and the model size you pick.

Try it in 5 minutes

A private API for your lab's scripts

Run `ollama run llama3`, then point a Python script at http://localhost:11434/v1 to summarize a results table and draft a figure caption.

Expected: A local LLM endpoint your scripts call like the OpenAI API — no data leaves the lab, no per-token bill.

Pricing

Free and open source. No subscription and no per-token cost — you only need a machine capable of running the model you choose.

Open the full Ollama course 5 lessons · cheat sheet

⚠️ Watch out

It is CLI- and API-first — for a graphical chat window pair it with a separate UI, or use LM Studio / Jan instead.

Cursor VS Code reimagined with an AI agent that writes and runs your code

Cursor

free tier cursor.com →

by Anysphere, Inc. · released 2023

VS Code reimagined with an AI agent that writes and runs your code

Cursor is a code editor built on VS Code — if you have ever used VS Code, the interface is immediately familiar, but AI is woven into every layer. As you type it autocompletes whole lines and functions; open the chat sidebar and you can ask questions about your own files in plain English; switch on Agent mode and Cursor reads your project, writes changes across multiple files, runs them, and iterates until it works. For a biology student this means you don't need Python syntax from memory: describe what you want — "load this CSV of gene-expression data, plot a heatmap, and flag any row with more than 20% missing values" — and Cursor writes the script. It understands pandas, NumPy, matplotlib, and Jupyter notebooks natively, and you can paste an error message straight into the chat to have it diagnosed and fixed. University students get a full year of Cursor Pro free via cursor.com/students.

Best for

Write code
Build an app

Also useful for

Data analytics
Automate workflow
Draft docs

What it does well

Built on VS Code — full extension ecosystem, familiar shortcuts, Jupyter notebook support out of the box
Agent mode autonomously reads your project, writes multi-file changes, runs the code, and fixes errors
Plain-English coding: describe what you want, no syntax knowledge required to get started
Supports all major frontier models (Claude, GPT, Gemini) — switch by task
Students with a .edu email get a full year of Pro free

Where it falls short

Free Hobby tier runs out of requests quickly; meaningful daily use needs the $20/month Pro plan
Sends code to cloud AI providers — not for confidential patient data or unpublished research without checking policy
Agent mode can drift on vague instructions; precise prompts and small task scopes work better
Heavier than a plain text editor — slower startup on older laptops

Try it in 5 minutes

Analyse a gene-expression CSV

I have a CSV called expression_data.csv with gene names in column A and values for 6 samples in B–G. Write a Python script that loads it with pandas, drops any gene with >20% missing values, and plots a clustered heatmap with seaborn.

Expected: Cursor generates a complete, ready-to-run Python script with imports, cleaning logic, and a seaborn clustermap.

Pricing

Hobby plan is free with limited monthly Agent requests and Tab completions (tight — expect to hit them in one focused session). Pro is $20/month ($16/month billed annually) with usage credits plus unlimited autocomplete; Pro+ $60/month; Ultra $200/month. University students can claim one free year of Pro at cursor.com/students with a .edu email.

Open the full Cursor course 7 lessons · cheat sheet

⚠️ Watch out

The free Hobby plan hits its limits fast; heavy use needs the $20/month Pro plan. Sends your code to AI providers — check institutional data policies before using on sensitive research data.

GitHub Copilot AI pair programmer built into your editor — writes code as you type

GitHub Copilot

free tier github.com/features/copilot →

by GitHub (Microsoft) · released 2022

AI pair programmer built into your editor — writes code as you type

GitHub Copilot, from GitHub (Microsoft) and OpenAI, sits directly inside your code editor. As you type it suggests the next line, the next function, or whole blocks — press Tab to accept or keep typing to ignore. Beyond autocomplete, a chat panel lets you ask "explain this error", "write a function that parses a FASTA file", or "convert this loop to a pandas one-liner", and an Agent Mode can plan, write, and run multi-step tasks across your project files. For biology students the most practical entry point is Python or R data analysis: open a .py or .R file in VS Code, install the free Copilot extension, and write a comment like "# load a CSV of gene-expression counts and plot a heatmap" — Copilot generates the code. It knows pandas, numpy, matplotlib, and Bioconductor workflows. Verified students can get unlimited completions free through Copilot Student, though new sign-ups via the Student Developer Pack are paused as of July 2026 — check before you count on it.

Best for

Write code
Automate workflow

Also useful for

Data analytics
Build an app
Draft docs

What it does well

Inline autocomplete is the fastest code-as-you-type experience — works in VS Code, JetBrains, Visual Studio, Xcode
Strong Python and R support — knows pandas, numpy, ggplot2, Bioconductor
Chat panel explains errors, suggests refactors, and answers "why does this code work?" in plain language
Agent Mode can write, run, and iterate across multiple files for longer tasks
Copilot Student gives verified .edu accounts free unlimited completions — when sign-ups are open (paused as of July 2026)

Where it falls short

Requires editor setup — no browser or standalone interface; install VS Code and the extension first
Can hallucinate plausible-but-wrong API calls and reference libraries that do not exist — always verify
Weaker on multi-file architectural changes and your specific biology-domain logic
Agent Mode can over-engineer or wander without explicit guidance

Try it in 5 minutes

Generate a gene-expression plot

# Load gene_counts.csv, normalize by library size, and plot a heatmap of the top 50 most variable genes with seaborn

Expected: Copilot generates ~15 lines of pandas + seaborn code — edit the filename and Tab through each suggestion.

Pricing

Free: up to 2,000 code completions/month plus limited chat, auto model selection only, no credit card. Pro: $10/month, unlimited completions and 1,500 AI credits. Pro+: $39/month, 7,000 credits. Max: $100/month, 20,000 credits. Business $19/seat/month; Enterprise $39/seat/month. Since June 2026 usage beyond completions is metered in GitHub AI Credits. Copilot Student is free for verified students with unlimited completions, but new sign-ups through the Student Developer Pack are paused as of July 2026 — verify before a cohort depends on it.

Open the full GitHub Copilot course 7 lessons · cheat sheet

⚠️ Watch out

Lives inside an IDE (VS Code, JetBrains, etc.) — you must install it first, a small setup barrier if you have never used a code editor.

Devin Desktop AI-native IDE whose agent codes whole tasks alongside you

Devin Desktop

free tier devin.ai/desktop →

by Cognition (formerly Codeium; renamed Windsurf to Devin Desktop in 2026) · released 2023

AI-native IDE whose agent codes whole tasks alongside you

Devin Desktop (formerly Windsurf) is an AI-native code editor (built on VS Code) whose centrepiece is Cascade — an autonomous agent that reads your whole project, plans changes across multiple files, runs terminal commands, installs dependencies, and iterates until the code works, without you describing each step. Unlike Copilot, which completes single lines, Cascade handles whole tasks: "add a CSV export button" or "refactor this analysis to use pandas". It stays aware of what you type and what runs in the terminal, so it can pick up mid-task. Windsurf began as Codeium and, after a much-publicised 2025 acquisition, was renamed Devin Desktop by its new owner Cognition (makers of the Devin agent) in 2026; existing installs updated over the air, with plans and settings carried over. For biology students with some Python, it can turn "parse this BLAST output and plot an e-value histogram" into working, runnable code in minutes.

Best for

Write code
Build an app

Also useful for

Automate workflow
Data analytics

What it does well

Cascade agent handles multi-file, multi-step tasks autonomously — describe a goal, not a function
Checkpoint snapshots let you revert any agent change in one click
Full VS Code extension ecosystem; minimal learning curve for anyone who has used VS Code
Tab autocomplete never counts against quota on any plan (including Free)
Built-in integrations connect to GitHub, databases, Slack, and Figma

Where it falls short

Free quota runs dry in 2–3 days of real use — heavy agentic sessions burn it fast
Requires programming knowledge; non-coders cannot meaningfully drive Cascade
Can spike CPU and slow down during long, complex agent runs
Recently renamed from Windsurf to Devin Desktop — some docs, tutorials, and screenshots elsewhere still use the old name

Try it in 5 minutes

Parse a BLAST result

I have a BLAST output file called blast_results.txt. Write a Python script that reads it, extracts the top hit per query (e-value, accession, description), and saves a clean CSV.

Expected: Cascade creates the script, explains the parsing logic, and runs it if you have Python installed.

Pricing

Free plan: unlimited Tab autocomplete plus a light Cascade quota (exhausted in ~2–3 real coding days). Pro: $20/month with access to Claude, GPT, and Gemini frontier models. Max: $200/month for heavy users. Annual billing saves ~17–20%.

Open the full Devin Desktop course 7 lessons · cheat sheet

⚠️ Watch out

Cognition renamed Windsurf to Devin Desktop in 2026 (windsurf.com redirects to devin.ai/desktop) — same editor, repositioned as an agent "command center". Needs coding basics, and the free Cascade quota runs out in a few days of real use.

Aider Open-source AI pair programmer in your terminal

Aider

API key required open source aider.chat →

by Paul Gauthier / Aider-AI (open source) · released 2023

Open-source AI pair programmer in your terminal

Aider is a free, open-source command-line tool that turns your terminal into an AI coding session. You point it at one or more files, describe what you want in plain English, and it edits those files and immediately commits the changes to Git with a descriptive message — so every AI change is a clean, reversible commit and nothing is silently overwritten. It supports 100+ programming languages, works with Claude, GPT, Gemini, DeepSeek, and locally-run models, and can map an entire codebase so the AI understands how files relate. Released in 2023 by Paul Gauthier, it has 40,000+ GitHub stars and a large, active community. For biology students: use it to write Python scripts that parse sequencing output, automate repetitive cleaning of CSV exports from lab instruments, or build a small analysis pipeline — with Git tracking every change so you can always roll back.

Best for

Write code
Open source / self-host

Also useful for

Build an app
Automate workflow
Data analytics

What it does well

Every AI edit is auto-committed to Git — changes are always reversible with /undo
Works with virtually any LLM: Claude, GPT, Gemini, DeepSeek, or local models via Ollama
Codebase mapping lets it reason across many files without you loading each one
Fully open-source (Apache 2.0) — no subscription, no vendor lock-in
Large, active community; a widely-cited public model-coding leaderboard

Where it falls short

Terminal-only — no GUI; requires basic command-line confidence to get started
You pay LLM API costs directly; a heavy session on a frontier model can add up
Less useful for pure writing tasks (lab reports, literature reviews) — it is a code editor
Windows users may want WSL2 for the smoothest experience

Try it in 5 minutes

Analyse a CSV in Python

aider --model sonnet analysis.py  →  then type: "Read data.csv and print mean ± SD for every numeric column"

Expected: Aider writes the script, runs it, and commits the change to Git — reversible with /undo.

Pricing

Aider itself is free and open-source (Apache 2.0). You pay only for the LLM API tokens you use — roughly $5–$15 per full day of active coding with Claude Sonnet; switching to a cheaper model like DeepSeek cuts that to a couple of dollars a day. It also runs against free local models via Ollama.

Open the full Aider course 6 lessons · cheat sheet

⚠️ Watch out

Requires comfort with the terminal and a paid LLM API key — not a point-and-click tool.

⚡ vLLM The serving engine built for many people hitting one model at once

⚡

vLLM

free open source vllm.ai →

by vLLM project — originated at UC Berkeley Sky Computing Lab (Kwon et al., PagedAttention, SOSP 2023), now a broad open-source community (2,000+ contributors) · released 2023

The serving engine built for many people hitting one model at once

vLLM is the engine you reach for when 'a model on my laptop' has to become 'a model the whole lab hits at the same time.' Built at UC Berkeley's Sky Computing Lab and now maintained by 2,000+ contributors, its core trick is PagedAttention — managing the GPU memory used for each conversation's attention cache the way an OS pages virtual memory, so many concurrent requests share a GPU without fragmenting it — combined with continuous batching and prefix caching. The result is dramatically higher throughput per GPU than a naive single-request server under real concurrent load. It speaks an OpenAI-compatible API (plus an Anthropic Messages API and gRPC), runs 200+ Hugging Face model architectures, and supports FP8/INT8/INT4/GPTQ/AWQ quantization to fit bigger models in less VRAM. For a lab: one GPU serving ten people's simultaneous questions instead of one.

Best for

Automate workflow
Data analytics

Also useful for

Open source / self-host

What it does well

PagedAttention + continuous batching — built specifically for many concurrent requests, not one
OpenAI-compatible + Anthropic Messages API + gRPC — drop into existing client code
200+ supported model architectures, broad quantization support (FP8/INT8/INT4/GPTQ/AWQ)
Runs on NVIDIA and AMD GPUs, plus TPU/Gaudi/Apple Silicon via hardware plugins

Where it falls short

GPU-first — CPU-only serving exists but is not where it shines (llama.cpp is the better fit there)
More moving parts to configure (tensor parallel, batching, quantization) than a single-user tool
VRAM sizing is a real planning exercise — under-provision and throughput (the whole point) suffers

Try it in 5 minutes

Serve one model to the whole lab at once

Run `vllm serve <a 7B-8B model>` on a 24GB+ GPU, then have three labmates hit its OpenAI-compatible endpoint with concurrent requests and compare wall-clock time against the same model in Ollama.

Expected: Noticeably steadier response times under concurrent load than a single-request server — the batching doing its job.

Pricing

Free and open source (Apache-2.0) — no license cost. The real cost is the GPU: a 7B–8B model needs roughly 8GB VRAM at 4-bit quantization just for weights, and meaningfully more (12–24GB+) for the context length and batch size a team actually needs — a single modern 24GB+ GPU is the realistic team-serving entry point (community heuristic, not a vLLM-published minimum — verify against your own model/context before buying hardware).

⚠️ Watch out

It is GPU-first and tuned for concurrent throughput, not the easiest single-user path — a solo laptop user is better served by Ollama or LM Studio.

🧱 llama.cpp The engine underneath Ollama — full control, runs on almost anything

🧱

llama.cpp

free open source github.com/ggml-org/llama.cpp →

by Georgi Gerganov + the ggml-org open-source community · released 2023

The engine underneath Ollama — full control, runs on almost anything

llama.cpp is the C/C++ engine that Ollama itself is built on — Ollama wraps it in a friendly Go layer (model pull, versioning, scheduling) that hides the knobs. This chapter teaches the layer underneath: llama.cpp's own server, for when a lab needs something Ollama's abstraction doesn't expose — a specific quantization, a build flag for an unusual CPU, or a single static binary with zero dependencies you can drop on hardware Ollama doesn't officially support. It reads the GGUF model format, and its stated goal is state-of-the-art performance on 'a wide range of hardware, locally and in the cloud': Apple Metal, x86 (AVX/AVX2/AVX512/AMX), RISC-V, NVIDIA CUDA, AMD HIP, Intel SYCL, Vulkan, and CPU+GPU hybrid — often the only realistic engine on a machine with no supported GPU at all. Its own server exposes the same OpenAI-compatible endpoints (chat, completions, embeddings), so it slots into this stack exactly where vLLM would, just with different trade-offs.

Best for

Automate workflow
Open source / self-host

Also useful for

Data analytics

What it does well

The actual engine under Ollama — this is what "using Ollama" is doing underneath, taught directly
Runs on almost anything: Apple Silicon, x86 with AVX/AMX, RISC-V, CUDA, HIP, SYCL, Vulkan, or CPU-only
Single static binary, minimal dependencies — starts serving in seconds on a fresh VM
OpenAI-compatible server with multimodal input, JSON-schema-constrained output, and LoRA adapters via `--lora`

Where it falls short

No accounts, UI, or multi-tenant story of its own — pair it with Open WebUI or LocalAI for a team-facing front end
Single-model-per-process by default — pair it with llama-swap to serve several models on one machine
The control it offers (build flags, quantization choices) is also its learning curve

Try it in 5 minutes

Run the engine underneath Ollama, directly

Download a GGUF model, start `llama-server` with it, and hit `POST /v1/chat/completions` with curl — the exact request Ollama sends internally, minus the wrapper.

Expected: The same kind of answer Ollama would give you, but now you can see (and change) every flag between the model and the response.

Pricing

Free and open source (MIT). No paid tier exists — cost is entirely the hardware you already have; its whole design point is running well on hardware that would not run vLLM at all (CPU-only, older GPUs, ARM, RISC-V).

⚠️ Watch out

It is a build-it-yourself engine, not a product — no accounts, no UI of its own, and the flags/build options that make it powerful also make it the most hands-on tool in this chapter.

🧩 LocalAI One OpenAI-compatible door in front of every engine you run

🧩

LocalAI

free open source localai.io →

by Ettore Di Giacinto + the LocalAI community · released 2023

One OpenAI-compatible door in front of every engine you run

LocalAI is the front door: one small core binary that speaks the OpenAI and Anthropic APIs, and routes each request to the right backend behind the scenes — llama.cpp for one model, vLLM for another, whisper.cpp for audio, Stable Diffusion for images — all through the same endpoint. Backends ship as OCI images and run as isolated gRPC processes, so a crash in one never takes down the others, and you install/upgrade a backend without touching the core. That's the concrete team win: n8n, AnythingLLM, or your own scripts all point at one stable URL forever, even as you swap which engine actually answers underneath — a new GPU, a new quantization, a different model family, none of it requires reconfiguring every downstream tool. It runs on NVIDIA, AMD, Intel, or Vulkan GPUs with autodetection, or CPU-only, and as of 2026 adds WebRTC real-time audio, MCP-based agents, and speaker diarization.

Best for

Automate workflow
Build an app

Also useful for

Open source / self-host
Data analytics

What it does well

One OpenAI + Anthropic-compatible endpoint in front of 35+ backends (llama.cpp, vLLM, Whisper, Stable Diffusion, MLX…)
Different models can run on different backends simultaneously behind the same URL
Backends are isolated processes — one crashing does not take the others down
Runs on NVIDIA/AMD/Intel/Vulkan GPUs or CPU-only, with autodetection

Where it falls short

Adds a hop — it is a router, so it inherits (never fixes) the throughput of whatever engine sits behind it
More moving parts than talking to one engine directly — worth it once you run more than one model/backend, overkill for just one
Smaller community than vLLM or llama.cpp alone, so some backend integrations lag the upstream project

Try it in 5 minutes

One URL, two different engines behind it

Configure LocalAI with two models — one served by llama.cpp, one by vLLM — and call both through the same `/v1/chat/completions` endpoint by changing only the `model` field.

Expected: Identical request shape for both models even though two completely different engines answered them.

Pricing

Free and open source (MIT) — no paid tier, no usage cap. Cost is the hardware behind each backend it manages, same as running that backend directly.

⚠️ Watch out

It's a router/gateway, not a faster engine — the throughput ceiling is still whichever backend (llama.cpp, vLLM…) it hands the request to.

🔁 llama-swap Hot-swap models on one GPU instead of keeping them all loaded

🔁

llama-swap

free github.com/mostlygeek/llama-swap →

by mostlygeek (GitHub handle; open-source, community-maintained) · released 2024

Hot-swap models on one GPU instead of keeping them all loaded

A lab GPU rarely has room to keep five models loaded at once — llama-swap is the small, single-purpose fix: a zero-dependency Go proxy that sits in front of llama.cpp, vLLM, or any OpenAI/Anthropic-compatible server, and starts, stops, or swaps the right backend process based on the `model` field in the incoming request. A YAML file maps each model name to the shell command that runs it; ask for a model that isn't loaded and llama-swap starts it (stopping another if VRAM is tight), with TTL-based auto-unload and 'groups' for the models you do want to keep resident together. It doesn't make any engine faster — it makes a single GPU behave like it's serving a whole model catalogue instead of one fixed model, which is exactly the gap between a personal Ollama box and a lab that wants to offer five different models without five GPUs.

Best for

Automate workflow

Also useful for

Open source / self-host

What it does well

Turns one GPU into "serves several models" instead of "serves one model" — no extra hardware
Works with llama.cpp, vLLM, and other OpenAI/Anthropic-compatible servers — not locked to one engine
Single static Go binary, zero dependencies, ships a small web UI for testing
TTL-based auto-unload and groups give explicit control over what stays resident

Where it falls short

Cold-start latency on the first request for a not-currently-loaded model
Adds another config file (the YAML model map) to keep in sync as models change
Solves a VRAM-sharing problem specifically — it is not a throughput or UI tool on its own

Try it in 5 minutes

Offer three models on one GPU

Write a llama-swap config mapping three different local models to three llama.cpp start commands, then request each by name in turn and watch it swap the running process.

Expected: One GPU answering for three different models, one at a time, without you manually stopping/starting anything.

Pricing

Free and open source (MIT) — no paid tier. Cost is unchanged: whatever GPU already runs your engine(s).

⚠️ Watch out

It swaps processes on demand, so the first request for a not-yet-loaded model pays a cold-start cost — not the right fit if every model must answer instantly, all the time.

No-code builders

Base44 No-code app builder with built-in database and hosting included

Base44

free tier base44.com →

by Base44 Inc. (founder Maor Shlomo; acquired by Wix, June 2025) · released 2024

No-code app builder with built-in database and hosting included

Base44 is an AI-powered no-code platform that generates fully functional web applications from plain English descriptions, including authentication, a built-in database, and hosting — all scaffolded automatically. Unlike tools that output only front-end code (like v0), Base44 produces a working backend with user roles, data persistence, and custom domain support out of the box. This makes it practical for building real internal tools, not just prototypes: a lab sample tracker, a PCR booking calendar, or a shared reagent inventory can go from prompt to live URL in under 10 minutes. Base44 was acquired by Wix in 2025 and is now Base44 Inc. under that umbrella. The free tier gives 25 message credits and 100 integration credits per month. Paid plans start at $16/month (annual) for 100 message credits and scale up to $160/month for 1,200. An AI "Superagent" feature can automate workflows inside your apps.

Best for

Build an app

Also useful for

Automate workflow

What it does well

Full-stack generation: creates both the UI and a working backend with user authentication, database tables, and role-based permissions — not just a prototype.
No deployment headaches: your app is hosted on Base44's infrastructure with a custom domain option from day one.
Integrations baked in: connects to Google Calendar, Slack, Notion, HubSpot, and Salesforce without custom API work.
Template marketplace: start from a pre-built "lab sample tracker" or "event scheduler" and customize from there.
Superagent feature: build AI-powered automations inside your app (e.g., auto-categorize incoming form submissions).

Where it falls short

Vendor lock-in: your data lives in Base44's infrastructure — migrating to your own server later requires significant re-work.
Message credits are consumed by every edit and refinement — complex apps with many iterations can exhaust the free tier quickly.
Not suited for performance-critical or HIPAA/GDPR-sensitive use cases without explicit vetting of Base44/Wix's compliance status.
Generated code is not designed to be exported and run independently — you're buying a managed platform, not portable code.

Try it in 5 minutes

Lab sample database in 5 minutes

Build a web app to track lab samples. Fields: date, sample ID, tissue type, storage location, notes. Include a sortable table and a CSV export button.

Expected: A working web app with form, table, and export — live URL provided, zero code written.

More prompts to try (1)

PCR booking calendar for the lab

Build a booking app for a shared thermocycler. Users can reserve 2-hour slots, see who has the machine booked, and get a confirmation email. Require login.

Pricing

Free tier: 25 message credits + 100 integration credits/month. Starter: $16/month (annual) for 100 message credits. Builder: $40/month. Pro: $80/month. Elite: $160/month for 1,200 message credits. All paid tiers include unlimited apps, GitHub integration, and custom domain support.

Open the full Base44 course 11 lessons · cheat sheet · recipes

⚠️ Watch out

Acquired by Wix in 2025 — generated apps include built-in auth and hosting but lock you into Base44/Wix for data storage and deployment.

🐳 Docker Containers, images, and volumes — the mental model before you install anything

🐳

Docker

free tier open source www.docker.com →

by Docker, Inc. · released 2013

Containers, images, and volumes — the mental model before you install anything

Docker packages an application together with everything it needs to run — code, runtime, system libraries — into a single reproducible unit called an image. Run that image and you get a container: an isolated, running instance of it, with its own filesystem and process space, but sharing the host's Linux kernel — which is why it starts in about a second instead of the minute-plus a full virtual machine needs to boot. A volume is Docker's answer to the one thing a container is deliberately bad at: keeping data. Containers are meant to be disposable, so anything that must survive a restart — a database, your workflow history — lives in a volume mounted into the container, not in the container's own writable layer. This is why nearly every local AI tool in this curriculum — Ollama, n8n, Open WebUI, AnythingLLM, Open Notebook — ships as a docker-compose.yml: the maintainer already solved "what Python version, what dependencies, what config" for you, and handed you one command to run it exactly as they tested it.

Best for

Automate workflow
Build an app

Also useful for

Write code
Data analytics

What it does well

One mental model — image (blueprint) → container (running instance) → volume (persistent data) — explains most of what confuses Docker beginners.
Reproducibility: the same image runs identically on your laptop, a teammate's machine, and a production server.
Composability: a docker-compose.yml wires multiple containers (app + database) together with one file and one command.
A container starts in about a second and shares the host kernel — much lighter than a full virtual machine.

Where it falls short

Containers are Linux processes under the hood — Windows and macOS both need a Linux VM layer underneath (WSL2 / Apple Hypervisor), which is where most install friction comes from.
A container's own filesystem is disposable by design — forgetting to mount a volume means your data vanishes on the next `docker compose down`.
Not everything needs Docker: a single script or a native desktop app (LM Studio) often runs fine, and faster, without the extra layer.

Try it in 5 minutes

See the image → container relationship for yourself

docker run hello-world && docker images && docker ps -a

Expected: Docker pulls the hello-world image (now visible in `docker images`), runs it once as a container that prints a message and exits — and `docker ps -a` still shows that finished container, separate from the image it came from.

Pricing

Docker Engine (the open-source core) is free everywhere. Docker Desktop — the GUI wrapper needed on Windows/macOS — is free for personal use, education, and companies under 250 employees / $10M revenue; larger organizations need a paid plan (roughly $9–24 per user per month).

Open the full Docker course 5 lessons

⚠️ Watch out

This chapter is concepts only, not an install guide — install Docker itself via the Windows, macOS, or Linux chapter once you know why you want it.

🐳 Docker on Windows Install Docker Desktop on Windows via WSL2 and run your first container

🐳

Docker on Windows

free tier open source www.docker.com/products/docker-desktop/ →

by Docker, Inc. · released 2013 (Docker); Docker Desktop for Windows 2016

Install Docker Desktop on Windows via WSL2 and run your first container

Docker lets you run software in containers — self-contained, reproducible boxes that behave the same on every machine, so "works on my computer" stops being a problem. On Windows the engine cannot run natively (containers are Linux), so Docker Desktop uses WSL2 — the Windows Subsystem for Linux 2, a genuine Linux kernel in a Hyper-V-managed VM — as its backend. This chapter takes you from nothing installed to a running container: enable and verify WSL2, install Docker Desktop, confirm the CLI can talk to the engine, then launch a real local n8n workflow with Docker Compose. It also demystifies the Windows → WSL2 → Docker-engine architecture and fixes the three pitfalls that trip most people up: virtualization disabled in BIOS, Hyper-V conflicts, and the WSL2 memory limit.

Best for

Automate workflow
Build an app

Also useful for

Write code
Data analytics

What it does well

Reproducible environments: a container runs identically on any machine, removing "works on my computer" bugs.
WSL2 backend gives near-native Linux container performance without leaving Windows.
Docker Compose spins up multi-service apps (app + database) from one YAML file.
Huge public image library (Docker Hub) — most tools ship a ready-to-run image.

Where it falls short

Most involved setup of the three OSes — needs WSL2 + BIOS virtualization enabled.
Docker Desktop is a background service that consumes RAM even when idle.
Hyper-V can conflict with other virtualization tools (some VirtualBox/VMware setups).

Try it in 5 minutes

Run your first container

docker run hello-world

Expected: Docker pulls a tiny test image and prints 'Hello from Docker!' — proof the engine works end to end.

Pricing

Docker Desktop is free for personal use, education, and small companies (under 250 employees and under $10M annual revenue). Larger organizations need a paid subscription (Pro/Team/Business, roughly $9–24 per user per month). The Docker engine and CLI themselves are open source.

Open the full Docker on Windows course 5 lessons

⚠️ Watch out

Windows is the most involved Docker setup because containers are Linux-only — Docker Desktop needs WSL2 (a real Linux kernel in a lightweight VM) underneath, plus hardware virtualization enabled in the BIOS.

🐳 Docker on macOS Install Docker Desktop on Mac (Apple Silicon or Intel) and run your first container

🐳

Docker on macOS

free tier open source www.docker.com/products/docker-desktop/ →

by Docker, Inc. · released 2013 (Docker); Docker Desktop for Mac 2016

Install Docker Desktop on Mac (Apple Silicon or Intel) and run your first container

Docker lets you run software in containers — self-contained, reproducible boxes that behave the same on every machine. On macOS, Docker Desktop runs a lightweight Linux VM via the Apple Hypervisor framework, so you get real Linux containers with a clean install. This chapter covers the install for both Apple Silicon (M1/M2/M3/M4) and Intel Macs, verifies the engine responds, and gets you running a real local n8n workflow with Docker Compose. It also covers what actually differs between Apple Silicon and Intel (image architecture and emulation), how to keep Docker's disk usage from ballooning, and the free-tier license terms so you know where the paid line sits.

Best for

Automate workflow
Build an app

Also useful for

Write code
Data analytics

What it does well

Cleanest of the three setups — one installer, no WSL2 or package-repo steps.
Native Apple Silicon support with fast arm64 images.
Docker Compose spins up multi-service apps from one YAML file.
Huge public image library (Docker Hub) — most tools ship a ready-to-run image.

Where it falls short

Apple Silicon runs Intel-only (amd64) images through emulation — slower, occasionally broken.
Docker Desktop is a background VM that consumes RAM and disk even when idle.
Image and build caches grow quietly; disk needs periodic pruning.

Try it in 5 minutes

Run your first container

docker run hello-world

Expected: Docker pulls a tiny test image and prints 'Hello from Docker!' — proof the engine works end to end.

Pricing

Docker Desktop is free for personal use, education, and small companies (under 250 employees and under $10M annual revenue). Larger organizations need a paid subscription (roughly $9–24 per user per month). The Docker engine and CLI are open source.

Open the full Docker on macOS course 5 lessons

⚠️ Watch out

The setup is straightforward, but on Apple Silicon some images are built only for Intel (amd64) and run through emulation — slower, and a few refuse to start without the right platform flag.

🐳 Docker on Linux Install Docker Engine natively on Linux, run without sudo, and start your first container

🐳

Docker on Linux

free open source docs.docker.com/engine/install/ →

by Docker, Inc. (Docker Engine, open source) · released 2013

Install Docker Engine natively on Linux, run without sudo, and start your first container

Docker lets you run software in containers — self-contained, reproducible boxes that behave the same on every machine. Linux is Docker's native home: you install the Docker Engine directly through your distribution's package manager, with no VM in between, so it is the most performant and lightweight of the three setups. This chapter covers Ubuntu/Debian (the most common) using the official Docker apt repository, shows you how to run Docker without sudo by joining the `docker` group, and gets you running a real local n8n workflow with Docker Compose. It also covers systemd integration so the engine starts on boot, and the Docker-socket security model so you understand what group membership actually grants.

Best for

Automate workflow
Build an app

Also useful for

Write code
Data analytics

What it does well

Native, no VM — the fastest and lightest Docker setup, lowest overhead.
Completely free and open source, no license tiers.
Installs from the official apt/dnf repository and integrates with systemd for auto-start.
Docker Compose spins up multi-service apps from one YAML file.

Where it falls short

Command-line only — no Docker Desktop GUI (though Desktop for Linux exists separately).
The `docker` group grants root-equivalent host access via the socket — a real security consideration.
Install steps differ across distributions (apt vs dnf vs pacman).

Try it in 5 minutes

Run your first container

docker run hello-world

Expected: Docker pulls a tiny test image and prints 'Hello from Docker!' — proof the engine works end to end.

Pricing

Docker Engine on Linux is fully free and open source — no license tiers, no paid subscription. (The paid Docker Desktop plans apply only to the Desktop app on Windows/macOS, not the Linux engine.)

Open the full Docker on Linux course 5 lessons

⚠️ Watch out

Adding your user to the `docker` group is what lets you skip sudo — but membership grants root-equivalent access to the host via the Docker socket, so only do it on a machine you trust.

⚖ Private AI for your org: buy, build & govern The data-residency spectrum + the governance checklist an institution asks

⚖

Private AI for your org: buy, build & govern

free open source gdpr.eu →

by Framework compiled from GDPR, the EU AI Act, EDPB guidance and CNIL recommendations — not a single vendor · released 2026

The data-residency spectrum + the governance checklist an institution asks

Private AI for your org is the decision framework a research team, an SME, or an institute — think EMBL Heidelberg — runs before letting AI touch real data. It places every option on one data-residency spectrum: a model running entirely on your own laptop, a self-hosted GPU server, an EU-managed SaaS workspace (Langdock, Mistral Le Chat Enterprise, Aleph Alpha PhariaAI), or a public API that may leave the EU — each step trading convenience for control. Layered on top is the governance checklist an institution actually runs: who is the data controller, is there a signed Data Processing Agreement (DPA/AVV), does the setup meet GDPR and the incoming EU AI Act, where does the data physically sit, and can access be audited after the fact. Worked buy-vs-build examples — an EU-managed SaaS against a self-hosted Open WebUI deployment — turn the checklist from theory into an actual procurement decision.

Best for

Open source / self-host
Automate workflow

Also useful for

Data analytics

What it does well

One spectrum to place any AI option on, from a laptop model to a public API — turns "is this compliant" into a repeatable question instead of a one-off debate.
A five-point governance checklist (controller, DPA, GDPR/EU AI Act fit, physical location, audit) that mirrors what an actual institutional data-protection office asks.
Grounded in real regulatory sources — GDPR, the EU AI Act, EDPB guidance and a CNIL recommendation — not vendor marketing.
Worked buy-vs-build examples with real numbers, so the trade-off is concrete rather than abstract.

Where it falls short

A framework, not a legal opinion — a real deployment still needs sign-off from your institution's actual data-protection officer.
The vendor landscape (Langdock, Mistral, Aleph Alpha, IONOS) moves fast; verify current certifications and pricing before you commit.
No single vendor has a published EMBL or Max Planck case study yet — the closest public precedent (Freiburg) is a university, not a research institute.

Try it in 5 minutes

Place your own AI use on the spectrum

Pick one real task you (or your team) currently do with AI. Write down: where does the model actually run, who is the data controller, and is there a signed DPA? If you cannot answer all three, it is not deployment-ready.

Expected: A one-paragraph answer that places your task on the data-residency spectrum and flags any missing item on the governance checklist.

More prompts to try (2)

Compare a buy path and a build path for one use case

Take a concrete task (e.g. "summarize incoming grant applications"). Sketch what it would take to run it via an EU-managed workspace (Langdock) versus self-hosting Open WebUI on a rented EU GPU server. List cost, setup time, and who controls the data for each.

Read one primary source and quote its actual recommendation

Read the CNIL recommendation on local AI deployment or the EDPB "AI Privacy Risks in LLMs" paper, and write two sentences on what it actually says about self-hosting vs. cloud AI — not what a secondary blog post claims it says.

Pricing

The framework itself is free to apply. Real cost depends on which rung of the spectrum you pick: self-hosting needs a GPU server (from a few hundred €/mo rented, more to own); an EU-managed workspace like Langdock runs roughly €25/user/month; a public API is pay-per-token with no seat fee but the weakest residency guarantee.

Open the full Private AI for your org: buy, build & govern course 3 lessons · recipes

⚠️ Watch out

A framework, not a legal opinion — a real deployment still needs sign-off from your own institution's data-protection officer, and the vendor landscape (Langdock, Mistral, Aleph Alpha, IONOS) moves fast, so verify current certifications before you commit.

🇩🇪 IONOS AI Model Hub German-hosted, OpenAI-compatible inference API — ISO 27001 data centers

🇩🇪

IONOS AI Model Hub

paid cloud.ionos.com/managed/ai-model-hub →

by IONOS SE (Germany) · released 2024

German-hosted, OpenAI-compatible inference API — ISO 27001 data centers

IONOS — Europe's largest hosting and cloud provider, headquartered in Germany — runs the AI Model Hub: a pay-as-you-go, OpenAI-compatible inference API serving open-weight models (Llama, Mistral and others) entirely from German, ISO 27001-certified data centers. For a German or EU institution the pitch is a straightforward answer to "where does the data sit": the same base-URL swap used for Groq or OpenRouter, but every request now stays on infrastructure governed by German law, with no US CLOUD Act exposure to explain to a data-protection officer. It sits at the EU-managed-SaaS rung of the data-residency spectrum — no hardware to operate, but a vendor and infrastructure both squarely inside the EU. It suits a lab or SME that wants hosted convenience without leaving GDPR's home turf, complementing rather than replacing a fully self-hosted setup for the most sensitive workloads.

Best for

Automate workflow
Open source / self-host

Also useful for

Data analytics

What it does well

Every request stays on German-operated infrastructure — a straightforward answer to "where does the data sit" for an EU institution.
ISO 27001-certified data centers, run by Europe's largest hosting provider.
OpenAI-compatible API — the same one-line base-URL swap used for Groq/OpenRouter/Cerebras.
No hardware to operate — the "buy" side of buy vs. build, without leaving EU jurisdiction.

Where it falls short

A smaller, evolving model catalog than the big US inference providers — check current model availability before committing.
A managed SaaS, not a self-hosted deployment — you still depend on a vendor's uptime and roadmap.
Less battle-tested at hyperscale than Groq/OpenRouter/Cerebras for very high-throughput workloads.

Try it in 5 minutes

Swap one base-URL to an EU-hosted endpoint

Point an existing OpenAI-compatible client at the IONOS AI Model Hub endpoint instead of Groq or OpenRouter, keep the request body identical, and compare the response.

Expected: The same completion, now served entirely from a German data center — nothing else in your code changes.

Pricing

Pay-as-you-go per-token pricing, billed in EUR, published on the IONOS Cloud pricing page — verify current rates and the supported model list before committing, as both change.

⚠️ Watch out

A smaller, evolving model catalog than the US hyperscalers — check ionos.com for the current list of supported open models before committing a workload.

🇫🇷 Scaleway French cloud — pay-per-token Generative APIs, or rent a GPU by the hour

🇫🇷

Scaleway

paid open source www.scaleway.com/en/generative-apis/ →

by Scaleway (Iliad Group, France) · released 2023

French cloud — pay-per-token Generative APIs, or rent a GPU by the hour

Scaleway is a French cloud provider (part of the Iliad group) offering two EU-sovereign paths to model inference. Generative APIs is a pay-per-token, OpenAI-compatible endpoint serving open models (Llama, Mistral, DeepSeek and others) from French data centers — the "buy" path, comparable to Groq or OpenRouter but EU-hosted throughout. GPU Instances is the "build" path: rent an H100-class GPU by the hour and run vLLM, Ollama, or any engine yourself — the same self-hosting playbook as a rented Hetzner box, just from a French provider under EU jurisdiction the whole way. Either way, Scaleway markets itself explicitly on European sovereignty and a lower carbon footprint than the US hyperscalers — relevant for an institution answering both a GDPR checklist and a sustainability one.

Best for

Automate workflow
Open source / self-host

Also useful for

Build an app

What it does well

Two paths under one EU-sovereign vendor: a hosted pay-per-token API, or a rented GPU you run yourself.
French data centers, French/EU jurisdiction — a clean answer to "where does this run" for a GDPR checklist.
OpenAI-compatible Generative APIs endpoint — the same base-URL swap pattern as Groq/OpenRouter.
Markets a lower carbon footprint than the US hyperscalers, relevant for a sustainability-conscious institution.

Where it falls short

Smaller model catalog and less raw throughput than Groq or Cerebras on the hosted-API side.
GPU Instances still require you to run and patch the serving stack yourself — not a managed inference API.
A smaller company than AWS/Azure/GCP — fewer regions and less enterprise tooling around the core product.

Try it in 5 minutes

Rent one GPU hour and serve a model yourself

Spin up a Scaleway GPU Instance, install Ollama or vLLM, load an open model, and call it from your own machine over its OpenAI-compatible endpoint — then shut the instance down.

Expected: An hour of real GPU inference on EU infrastructure, billed only for the time it ran — the same pattern as self-hosting on a rented Hetzner box.

Pricing

Generative APIs: pay-per-token, published on scaleway.com. GPU Instances: pay-per-hour by GPU class, typically below equivalent AWS/Azure/GCP rates — verify current pricing before budgeting, as both products update independently.

⚠️ Watch out

Two different products under one name — hosted Generative APIs (pay-per-token, zero ops) and raw GPU Instances (pay-per-hour, you run the stack) — pick based on how much control you actually need.

🇫🇷 OVHcloud Europe's largest independent cloud — serverless AI Endpoints, EU-hosted

🇫🇷

OVHcloud

paid www.ovhcloud.com/en/public-cloud/ai-endpoints/ →

by OVH Groupe SA (France) · released 2023

Europe's largest independent cloud — serverless AI Endpoints, EU-hosted

OVHcloud is a French, publicly listed cloud infrastructure company — one of the largest cloud providers headquartered in Europe — and its AI Endpoints product is a serverless, pay-per-use inference API for open-weight models (Llama, Mistral and others), hosted entirely in OVHcloud's own European data centers. Like IONOS and Scaleway, it targets the same OpenAI-compatible base-URL swap as Groq or OpenRouter, but every request stays under EU/French jurisdiction with no dependency on a US hyperscaler. OVHcloud also rents bare GPU instances for teams that want to run their own serving stack (vLLM, Ollama) rather than a managed endpoint. For a research institute, the appeal is that OVHcloud is a large, independently audited European company with a two-decade infrastructure track record — not a startup betting its first funding round on GDPR compliance.

Best for

Automate workflow
Open source / self-host

Also useful for

Data analytics

What it does well

A large, independently audited, publicly listed European cloud company — a two-decade infrastructure track record, not a startup.
Serverless AI Endpoints: pay-per-use, OpenAI-compatible, zero ops — the "buy" path, fully EU-hosted.
Also rents raw GPU instances for a "build" path under the same EU vendor.
French/EU jurisdiction throughout — no US CLOUD Act exposure to explain to a data-protection officer.

Where it falls short

Smaller open-model catalog on AI Endpoints than Groq/OpenRouter's selection.
Historically stronger on general cloud infrastructure than on cutting-edge AI tooling.
Raw GPU instances still need you to run and maintain the serving stack yourself.

Try it in 5 minutes

Call an EU-hosted model with your existing OpenAI client

Point your existing OpenAI-compatible code at an OVHcloud AI Endpoints URL instead of a US provider, keep the request body identical, and confirm the response.

Expected: The same completion, now served entirely from an OVHcloud European data center — only the jurisdiction changed.

Pricing

Pay-per-token on AI Endpoints, or pay-per-hour on rented GPU instances — both published on ovhcloud.com; verify current rates, which are typically below equivalent US-hyperscaler pricing.

⚠️ Watch out

Historically stronger on general cloud infrastructure than on cutting-edge AI tooling — expect a narrower AI-specific feature set than a dedicated inference specialist.

Lovable Chat to build full-stack React apps with GitHub export

Lovable

free tier lovable.dev →

by Lovable (Anton Osika, CEO; Fabian Hedin, co-founder; Stockholm) · released Nov 2024 (as Lovable; GPT Engineer App launched 2023)

Chat to build full-stack React apps with GitHub export

Lovable was founded in Stockholm in 2023 by Anton Osika and Fabian Hedin under the name "GPT Engineer App," rebranded to Lovable in December 2024, and achieved unicorn status ($1.8 billion valuation) in July 2025 after raising $200 million Series A. By December 2025 it had raised an additional $330 million at a $6.6 billion valuation and was approaching 8 million users. The platform lets you describe an application in chat, watch a live preview update in real time, and push the generated code to a GitHub repository with one click. Technically it generates React + TypeScript + Tailwind CSS applications with Supabase as the backend (authentication and database). You can refine the app iteratively — connect Figma designs, describe UI tweaks, or add integrations like Stripe and Resend. For biology students, Lovable shines at building shareable tools: a lab results visualizer, a gene expression dashboard, or a booking form — and the resulting code is portable.

Best for

Build an app

What it does well

Live preview updates as you chat — you see the UI change in real time while you describe what you want.
GitHub export in one click: the full React/TypeScript/Tailwind codebase is yours to own, run, and deploy independently.
Supabase integration auto-wires authentication and database tables — no manual schema design needed for standard use cases.
Figma import: paste a design or screenshot and Lovable matches the layout, which is useful for replicating an existing paper figure as an interactive chart.
Mobile app (iOS + Android) for continuing builds on the go.

Where it falls short

Credit-based model with unpredictable consumption: iterative debugging loops (fixing a layout bug across 5 chat messages) can drain credits 10x faster than expected.
Backend flexibility is limited to Supabase — if you need a different database or a Python backend, Lovable is not the right tool.
AI-generated code sometimes introduces subtle logic errors that are hard to debug if you don't read TypeScript.
Complex apps with many features require many refinement cycles — expect iterative generation rather than one-shot perfection.

Try it in 5 minutes

Sample log form with CSV export

Build a one-page form to log lab samples (date, sample ID, tissue type, notes). Add a table view of logged entries and an "Export as CSV" button.

Expected: Visual React app preview in-browser; push to GitHub in one click.

More prompts to try (1)

Gene expression bar chart dashboard

Build a dashboard with a file upload for a CSV (columns: gene, condition_A, condition_B). Display a grouped bar chart comparing expression values per gene and highlight any gene with >2-fold change in red.

Pricing

Free: 5 credits/day (reset midnight UTC, max ~30/month, no rollover). Pro: $25/month for 100 monthly credits (these roll over) plus the 5 daily credits — capped at 150/month — with custom domains, badge removal, user roles and Code-mode code editing; the monthly allowance scales up to 10,000 credits. Business: $50/month (100 credits) adds SSO, restricted (invite-only) projects, data-training opt-out and reusable design templates. Every workspace — including Free — also currently gets a free $25 Cloud + $1 AI usage balance each month for hosting and built-in AI (Lovable lists this as a temporary offering). Student discount up to 50% off Pro — verify at lovable.dev/pricing.

Open the full Lovable course 10 lessons · cheat sheet · recipes

⚠️ Watch out

Credit consumption can surprise you — complex features cost 1–2 credits but debugging loops can burn 50–150 credits before a stubborn bug is fixed.

v0 Prompt full-stack React apps — live preview, one-click deploy on Vercel

v0

free tier v0.app →

by Vercel (Guillermo Rauch, CEO) · released 2023; relaunched as a full agent Feb 2026

Prompt full-stack React apps — live preview, one-click deploy on Vercel

v0 is Vercel's AI agent for building real code and full-stack apps. From a natural-language prompt it generates React + Next.js using shadcn/ui components and Tailwind CSS, shows a production-accurate live preview, and deploys to Vercel in one click. A major relaunch in February 2026 turned it into a full agent: it can search the web (with inline source links), use a browser to test and critique the apps it builds, run terminal commands in an isolated sandbox, fix its own errors, import existing GitHub repositories, and edit code in a built-in VS Code-style editor. It wires up one-click databases (Supabase, Neon, Upstash, Vercel Blob) and AI models (through the Vercel AI Gateway, no API key required) so apps can be genuinely data-driven — though it stays a React/Next.js tool at heart, strongest on polished front-ends. You own the output: every project syncs to GitHub as standard Next.js source you can clone and run anywhere.

Best for

Build an app

Also useful for

Write code

What it does well

Generates polished, accessible UI from shadcn/ui + Tailwind CSS — professional output with no design work.
A full agent since Feb 2026: it searches the web, uses a browser to test and critique what it builds, runs terminal commands in a sandbox, and fixes its own errors.
One-click databases (Supabase, Neon, Upstash, Vercel Blob) and AI models (via the Vercel AI Gateway) make apps genuinely full-stack.
Design Mode: select an element and tweak its type, colour, spacing and layout visually — the edits apply back to your code.
Production-accurate previews and one-click Vercel deploy; every project syncs to GitHub as standard Next.js code you own.

Where it falls short

It is a React/Next.js + TypeScript tool — the output assumes that stack, so a plain-HTML, Vue or Python project needs heavy adaptation.
Free tier is ~7 messages/day plus $5 of monthly credits — a single iterative session can exhaust it.
Backends are "bring a provider" integrations (Supabase, Neon, Upstash, Vercel Blob), not one fully managed built-in cloud — you create and manage those accounts.
Token-metered: the v0 Max and Max Fast models get expensive on the API ($25–$50 per 1M output tokens) — pick the smallest model that does the job.

Try it in 5 minutes

Protein structure data table with filters

Build a responsive data table for protein structure data. Columns: protein name, PDB ID, resolution (Å), method (X-ray/cryo-EM/NMR), organism. Add a dropdown filter by method.

Expected: React component with live preview; copy code or deploy to Vercel in one click.

More prompts to try (1)

Interactive gene expression heatmap widget

Build a React component that accepts a 2D array of gene expression values and renders an interactive color-coded heatmap. Clicking a cell shows the gene name and expression value in a tooltip.

Pricing

Free: $0/month — $5 of included monthly credits and a 7-messages/day limit; includes deploying to Vercel, Design Mode and GitHub sync. Team: $30/user/month ($30 monthly credits/user + $2/day login credits, shared team credits, collaboration). Business: $100/user/month (same credits, plus training opt-out by default, SSO and role-based access). Enterprise: custom. Usage is metered by tokens against the model you pick — v0 Mini ($1 in / $5 out per 1M tokens), v0 Pro ($3 / $15), v0 Max ($5 / $25), v0 Max Fast ($10 / $50). Students get 1 year of free v0 premium if their school is on the active list (Stanford, MIT, UC Berkeley, NYU and more) — check v0.app/students. Verify current numbers at v0.app/pricing.

Open the full v0 course 10 lessons · cheat sheet · recipes

⚠️ Watch out

v0 is a React/Next.js tool. Since its Feb 2026 relaunch it builds full-stack apps via one-click database and API integrations — but a non-JavaScript stack or a heavy custom backend still sits outside its sweet spot.

Make Visual no-code workflow automation

Make

free tier make.com →

by Make (a Celonis business unit) · released 2016

Visual no-code workflow automation

Make (formerly Integromat) is a visual drag-and-drop platform for connecting apps and automating multi-step workflows — no code required. You design "scenarios": flowchart-like canvases where data travels from a trigger (a new row in Google Sheets, an incoming email) through a chain of actions (send to Slack, update Airtable, call an API, run an AI model). With 3,000+ pre-built app integrations — including Google Workspace, Notion, PubMed via HTTP, Slack, and 400+ AI services — Make covers the connective tissue of a modern lab without writing a single line of code. Its key differentiator over Zapier is a fully visual canvas with branching logic, loops, error handlers, and data transformation built right into the diagram. For biology researchers, practical starting points include automating literature-alert emails into a shared database, triggering data-export pipelines when an instrument writes a file, or routing student form submissions to grading spreadsheets automatically.

Best for

Automate workflow
Build an app

Also useful for

Templates
Data analytics

What it does well

Fully visual canvas with branches, loops, and error-handling — more expressive than linear Zapier steps
3,000+ pre-built integrations including Google Workspace, Notion, Slack, OpenAI, and HTTP/webhook for any API
Generous free tier requires no credit card and covers simple personal automations
Real-time execution view shows data flowing through each node — easy to debug
AI-assisted scenario builder (Maia) lets you describe a workflow in plain language

Where it falls short

Free tier: only 2 active scenarios and 1,000 ops/month — exhausted quickly by any real pipeline
Learning curve is steeper than Zapier; the canvas can feel overwhelming for first-timers
No built-in data storage — you must connect an external database (Airtable, Google Sheets, Supabase)
Complex scenarios with many branches can become visually cluttered and hard to maintain

Try it in 5 minutes

Auto-file PubMed alerts into a spreadsheet

Create a scenario: trigger on a new email from pubmed-no-reply@ncbi.nlm.nih.gov → parse the subject line → append the title + date as a Google Sheets row.

Expected: Each new PubMed email alert automatically adds a row to your literature-tracking sheet within minutes of arrival.

Pricing

Free: $0/month — 1,000 operations, 2 active scenarios, 15-min minimum trigger interval. Core: $9/month, Pro: $16/month, Teams: $29/month — all start at 10,000 operations/month with 1-min intervals; annual billing saves ~15%. Enterprise: custom.

⚠️ Watch out

The free tier caps you at 1,000 operations/month and only 2 active scenarios, so any real research pipeline will quickly need a paid plan.

Zapier Connect 9,000+ apps, no code needed

Zapier

free tier zapier.com →

by Zapier Inc. · released 2011

Connect 9,000+ apps, no code needed

Zapier is the most widely adopted no-code automation platform, used by 3 million+ businesses to link 9,000+ apps via point-and-click "Zaps" — each Zap being a trigger → action chain that fires automatically. No programming knowledge is required: pick a trigger app (Google Forms, Gmail, Slack), choose an event, then chain one or more actions in other apps. Multi-step Zaps, conditional paths (Filters), and data formatting are available on paid plans. Since 2024–2025 Zapier has added AI-native features: a Copilot builder, AI Agents, Zapier Tables (a lightweight database), and MCP integration so tools like Claude can trigger automations directly. For biology researchers and lab managers, practical entry points are automatic logging of survey or form responses into Google Sheets, routing email alerts when instrument data arrives, syncing reference managers, or linking Slack/email to a task tracker — all without touching code.

Best for

Automate workflow
Templates

Also useful for

Build an app
Data analytics

What it does well

9,000+ app integrations — by far the largest catalog of any no-code automation platform
No-code visual editor: lab staff with zero programming background can build working automations in minutes
Huge template library — hundreds of ready-made Zaps for common admin tasks (form → sheet, email → task)
AI Copilot generates Zap drafts from plain-English descriptions, lowering the setup barrier further
Reliable, mature infrastructure with detailed run logs for debugging

Where it falls short

Task-based pricing gets expensive fast: multi-step Zaps consume one task per action step per run
Free plan is very limited (100 tasks/month, two-step only, 15-min polling) — inadequate for most real workflows
Not suitable for high-frequency or real-time pipelines; minimum polling is 1–2 min on paid plans
Complex branching and data transformation are awkward compared to code-based tools like n8n

Try it in 5 minutes

Log a Google Form response to Sheets

Create a Zap: Trigger = "Google Forms – New Response", Action = "Google Sheets – Create Spreadsheet Row". Map each form field to a column, then turn it on.

Expected: Every new form submission automatically appears as a new row in your sheet — no manual copy-paste.

Pricing

Free: 100 tasks/month, two-step Zaps only. Professional: $29.99/mo (or $19.99/mo billed annually) for 750 tasks/month, multi-step Zaps, and premium apps. Team: $103.50/mo (or $69.50/mo annually) for 2,000 tasks and up to 25 users. Enterprise: custom.

⚠️ Watch out

Task-based billing escalates sharply — a 5-step Zap counts 5 tasks per run, so the free tier (100 tasks/month) disappears fast on any real workflow.

KNIME Visual data science — no code needed

KNIME

free open source www.knime.com →

by KNIME AG (founded at the University of Konstanz) · released 2006

Visual data science — no code needed

KNIME Analytics Platform is a free, open-source visual workbench where you build data pipelines by connecting "nodes" on a canvas — no programming required. Each node performs one step: read a CSV, filter rows, run a statistical test, plot a chart, export results. You chain them into a workflow that runs top-to-bottom, reproducibly, every time. With 300+ connectors (Excel, databases, cloud storage, REST APIs) and built-in machine learning, it covers the full data-science lifecycle from raw data to publication-ready output. A growing AI/LLM extension lets you wire in language-model nodes for text classification, summarisation, and RAG pipelines — all still drag-and-drop. For biology researchers at COS Heidelberg, KNIME is ideal for cleaning messy instrument exports (mass-spec, NGS, flow-cytometry CSVs), building a reproducible no-code analysis pipeline, and sharing that exact workflow as a "recipe" with collaborators — no Python environment needed.

Best for

Data analytics
Automate workflow

Also useful for

Build an app
Learning resources

What it does well

Genuinely no-code: drag-and-drop node canvas with 300+ connectors, readable by anyone
Strong life-sciences pedigree — used in drug discovery, NGS, and clinical data pipelines
AI/LLM extension: plug in OpenAI, Ollama, or local models as ordinary nodes
Open-source desktop is fully capable; sharing workflows as portable "recipes" is free
Active community hub with thousands of pre-built workflow templates

Where it falls short

Performance degrades on very large datasets (>1 GB in-memory); not a Spark/Dask replacement out of the box
Visual canvas becomes hard to read with complex multi-branch workflows (node sprawl)
Team sharing, scheduling, and web execution require a paid Hub plan
Initial setup and node discovery has a learning curve despite the no-code promise

Try it in 5 minutes

Import & clean a spreadsheet

Install KNIME Analytics Platform. Drag a "CSV Reader" node onto the canvas, point it at a lab CSV, add a "Row Filter" node, connect them, and hit the green play button.

Expected: A live data table showing your filtered rows — with no code written.

Pricing

The desktop KNIME Analytics Platform is fully free and open-source. The cloud collaboration layer — KNIME Community Hub — adds a paid Pro tier (~$19/month) and Team tier (~$99/month for up to 3 users); enterprise Business Hub pricing is on request.

Open the full KNIME course 5 lessons · cheat sheet

⚠️ Watch out

Memory-hungry: large datasets (>1 GB) can slow or crash the desktop app, and team collaboration/scheduling needs a paid Hub subscription.

Flowise Build AI agents visually — no code

Flowise

free tier open source flowiseai.com →

by FlowiseAI — Henry Heng & Chung Yau Ong (Y Combinator); acquired by Workday Aug 2025 · released 2023

Build AI agents visually — no code

Flowise is an open-source, visual tool for building AI agents and chatbots without writing code. Instead of programming, you drag boxes called "nodes" onto a canvas and connect them — one node for a language model, one for your documents, one for a search tool — so you literally see how information flows. A non-programmer can assemble a working assistant: e.g. a chatbot that answers questions from your own PDFs of lab protocols or papers (a technique called RAG, retrieval-augmented generation). You build and test in the browser, then publish the agent as a chat widget you embed on a site or call from other software. It sits between hand-written code and rigid pre-made apps — more flexible than a fixed chatbot builder, far gentler than coding an agent from scratch. Supports 100+ models, vector databases, and tools.

Best for

Build an app
Automate workflow

Also useful for

Open source / self-host
Find papers

What it does well

Visual drag-and-drop canvas — build working AI agents with no coding
100+ integrations: major LLMs, embeddings, and vector databases
Strong for RAG — turn your own PDFs/notes into a Q&A chatbot
Single chat assistants or multi-agent workflows, with human-in-the-loop review
Self-host for free, or publish agents as embeddable widgets / APIs (TS & Python SDKs)

Where it falls short

You pay separately for LLM API usage — the builder is free, the models usually are not
Concepts like vector stores, embeddings, and chains have a real learning curve
Self-hosting means installing/running software (Docker or Node.js)
Direction is less certain after the 2025 Workday acquisition, though it stays open-source

Try it in 5 minutes

Build a PDF question-answering chatbot

Start a free Flowise Cloud account → New Chatflow → drag in a "PDF File" loader and upload one paper → connect an OpenAI chat node + a vector store → open the chat panel and ask: "Summarize the main method in this document."

Expected: A chat reply grounded in the PDF you uploaded — not generic knowledge.

Pricing

Self-hosting the open-source version (Apache 2.0) is free. Flowise Cloud has a free plan (2 flows, 100 predictions/mo) with paid managed plans reportedly from ~$35/mo — re-check the live pricing page. You also pay your own LLM API usage.

Open the full Flowise course 7 lessons · cheat sheet

⚠️ Watch out

The visual builder is free, but you bring (and pay for) your own LLM API key; self-hosting needs Docker or Node.js.

Langflow Drag-and-drop builder for AI agents & RAG

Langflow

free open source www.langflow.org →

by Logspace (Rodrigo Nader) → DataStax → IBM (open source) · released 2023

Drag-and-drop builder for AI agents & RAG

Langflow is a free, open-source visual tool for building AI applications without writing much code. You drag pre-made blocks — an LLM, a document store, a chat input, an agent, a web-fetch tool — onto a canvas and wire them to design how information flows. It's a friendly front end over Python and LangChain: every block is real Python you can open and edit, but beginners never have to. A built-in Playground lets you chat with your flow and watch the agent reason and pick tools in real time. Typical builds are RAG chatbots that answer questions from your own PDFs or a website, and multi-step agents. Finished flows export as an API or MCP server, so a scientist could turn a literature-Q&A bot or a data-lookup assistant into something a whole lab can use.

Best for

Build an app
Open source / self-host

Also useful for

Automate workflow
Find papers

What it does well

Visual drag-and-drop canvas — assemble agents and RAG pipelines with little or no coding
Free, MIT-licensed; runs fully on your own laptop or server — no vendor lock-in
Model-agnostic: works with major LLMs and vector databases, bring-your-own keys
Live Playground shows the agent's step-by-step reasoning and tool choices
One-click export of any flow as a REST API or MCP server; every block is editable Python

Where it falls short

No first-party managed cloud anymore — you self-host or use a third-party host
Self-hosting still needs setup (install, API keys) and a machine to run it
Long-term direction now sits with IBM/watsonx
Many components can feel overwhelming; complex flows benefit from light Python knowledge

Try it in 5 minutes

Run the Simple Agent template

Open Langflow → New Flow → pick the "Simple Agent" template → paste an OpenAI API key → click "Playground" → type: "I want to add 4 and 4."

Expected: The agent shows its reasoning, picks the Calculator tool, and answers 8 — you watch it think.

Pricing

Free and open-source (MIT) to download and self-host — you pay only your own infrastructure plus any LLM API keys you bring. The former DataStax-hosted cloud was retired April 2026; managed hosting now goes via partners (IBM watsonx, Render).

Open the full Langflow course 8 lessons · cheat sheet

⚠️ Watch out

Free and open-source, but you self-host (the hosted DataStax cloud closed Apr 2026) and bring your own model API key.

CrewAI Python framework for teams of AI agents

CrewAI

free tier open source www.crewai.com →

by CrewAI Inc. — João (Joe) Moura · released 2023

Python framework for teams of AI agents

CrewAI is an open-source Python framework for building multi-agent AI systems. You define several role-playing agents (e.g. a "researcher" and a "writer"), give each a goal and tools, and assign them tasks; together they form a "crew" that collaborates to complete complex work. It also offers "Flows" for more structured, event-driven control. Built from scratch with no dependency on LangChain, it emphasizes a lean, fast runtime. Crucially, it is code-first: you write Python (configured via YAML + Python files) rather than dragging boxes on a canvas, so it suits learners comfortable with a little scripting. It sits a step beyond visual no-code tools like Flowise or n8n — more flexible and powerful, but you trade the visual builder for actual code. A paid hosted platform adds a visual editor, deployment, and monitoring on top.

Best for

Write code
Build an app

Also useful for

Open source / self-host
Automate workflow

What it does well

Multiple specialized agents collaborate on one task ("crew") — models real division of labor
Lean and standalone — built from scratch, independent of LangChain
Open-source (MIT) with a very large community (50k+ GitHub stars)
Two paradigms: autonomous "Crews" plus structured event-driven "Flows"
Model-agnostic — many LLM providers, plus a large library of tools/integrations

Where it falls short

Code-first: you must write Python (and YAML config) — not a no-code/visual tool
Multi-agent orchestration adds conceptual overhead vs a single prompt or chatbot
You bring (and pay for) your own LLM API key for the framework to run
Advanced deploy/monitor/visual-editor features sit behind the hosted platform

Try it in 5 minutes

Scaffold your first crew

In a terminal with Python 3.10+: `pip install crewai` → `crewai create crew my_first_crew` → add your LLM API key to the generated .env → `crewai run`

Expected: Two agents (a researcher and a writer) print their reasoning and produce a short written report.

Pricing

The core Python framework is free and open-source (MIT). A hosted platform has a free Basic tier (visual editor + 50 workflow runs/mo) and a quote-only Enterprise tier. You bring your own LLM API key to actually run agents.

Open the full CrewAI course 9 lessons · cheat sheet

⚠️ Watch out

Code-first: you write Python (not drag boxes) — best if you are comfortable with a bit of scripting.

🎙️ ElevenLabs The most human-sounding voice AI, plus cloning and dubbing

🎙️

ElevenLabs

free tier elevenlabs.io →

by ElevenLabs · released 2022

The most human-sounding voice AI, plus cloning and dubbing

ElevenLabs is the quality benchmark for text-to-speech: paste in text and get back audio that is very hard to distinguish from a real person, including emotional tone and pacing. Its two standout features for a student's project are voice cloning (record a short sample, get a synthetic version of that voice reading anything) and dubbing (translate and re-voice a video into another language automatically). Latency is low enough (roughly 75-300ms) that it can be used in live, conversational apps, not just pre-rendered narration. Free accounts get 10,000 credits a month — enough to try it and make a short clip, but not enough for a real project. This is the tool to reach for when the output needs to sound convincingly human, e.g. narrating a lab-safety video or building a voice demo, rather than just being intelligible.

Best for

Build an app

Also useful for

Draft docs

What it does well

Most natural-sounding voices of any TTS provider, cloud or local
Voice cloning from a short audio sample
Automatic dubbing/translation of video audio
Low enough latency (~75-300ms) for live conversation use

Where it falls short

Free tier is a taste, not a workflow — real use needs a paid plan
Cloud-only: your text and any cloned voice sample leave your machine
Cloning a real person's voice raises consent questions worth thinking about first

Try it in 5 minutes

Clone your own voice and read a paragraph

On elevenlabs.io, record ~1 minute of yourself reading, create an Instant Voice Clone, then paste a paragraph and generate audio in your own cloned voice.

Expected: A 30-60 second clip that sounds unmistakably like you, generated in about 10-20 seconds.

Pricing

Free tier: 10k credits/mo. Creator: $22/mo (~100k credits). Pro: $99/mo. API pay-as-you-go roughly $0.05-0.10 per 1,000 characters; dubbing $0.33-0.50/min.

⚠️ Watch out

The good voices and cloning live behind the paid tiers — free credits run out fast.

⚡ Cartesia (Sonic) Sub-100ms streaming TTS built for real-time voice agents

⚡

Cartesia (Sonic)

API key required cartesia.ai →

by Cartesia · released Sonic-2: February 2026

Sub-100ms streaming TTS built for real-time voice agents

Cartesia's Sonic model is built for one thing: speed. It streams audio back in well under 100ms, which is the difference between a voice agent that feels like a real conversation and one that feels like a phone menu with an awkward pause after every sentence. Sonic-2, released February 2026, covers 15+ languages and is positioned specifically as the latency leader among TTS APIs — competitors like ElevenLabs are excellent but noticeably slower to start speaking. For a student this is less "something you click a button on" and more the engine behind a live voice assistant, an interactive lab tutor, or any app where the AI needs to start talking back almost instantly. It is priced to be cheap at volume, which matters if you are building something that talks a lot rather than narrating a single video.

Best for

Build an app

Also useful for

Talk through ideas

What it does well

Fastest time-to-first-audio of any major TTS API (sub-100ms)
15+ languages in Sonic-2
Built specifically for real-time, streaming voice agents
Cheap enough to run at high volume

Where it falls short

No polished consumer app — you need to call the API from code
Voice quality is very good but not always rated above ElevenLabs
Cloud-only, requires an API key and billing setup

Try it in 5 minutes

Read the streaming quickstart

Open the Cartesia docs quickstart and run the sample streaming request with a free trial key, timing how quickly the first audio chunk arrives.

Expected: First audio arriving in well under a second — noticeably faster than pasting the same text into a typical TTS website.

Pricing

Usage-based API pricing, positioned as inexpensive at volume (roughly cents per million characters); requires your own API key, no consumer web app.

⚠️ Watch out

It is a developer API, not a website you paste text into — you need someone to wire it up.

🔊 OpenAI TTS The cheapest good cloud voice, bundled with the GPT ecosystem

🔊

OpenAI TTS

API key required platform.openai.com/docs/guides/text-to-speech →

by OpenAI · released tts-1: Nov 2023

The cheapest good cloud voice, bundled with the GPT ecosystem

OpenAI's TTS models (tts-1 for speed, tts-1-hd for quality) turn text into speech at a price that undercuts almost every competitor — around $0.015 per million characters for tts-1 — which matters if you want to narrate a lot of content, like turning every lecture slide deck into audio. It plugs neatly into anything already using ChatGPT/GPT-4o, so if you're already scripting content with GPT you can pipe the output straight to speech without adding a new vendor. OpenAI also offers gpt-4o-transcribe on the speech-to-text side, so the same account can handle both directions of a voice pipeline. It won't win a blind listening test against ElevenLabs, but for straightforward narration — audio versions of readings, a voice for a chatbot, accessibility read-aloud — it's the practical default because it's cheap and simple.

Best for

Build an app

Also useful for

Draft docs

What it does well

Cheapest well-known cloud TTS by a wide margin
Same API key/account as ChatGPT and GPT-4o — one vendor for text and voice
Also offers gpt-4o-transcribe for speech-to-text
Simple, well-documented API

Where it falls short

Voice quality and expressiveness trail ElevenLabs and Cartesia
No voice cloning
Cloud-only, requires an API key

Try it in 5 minutes

Narrate a short abstract

Call the OpenAI TTS API (tts-1) with a paragraph from a paper abstract and listen to the output; compare the cost shown against ElevenLabs.

Expected: Clear, understandable narration for a fraction of a cent, generated in a couple of seconds.

Pricing

tts-1 (fast): ~$0.015 per 1M characters. tts-1-hd: higher quality, higher price. Billed via OpenAI API usage, no separate subscription.

⚠️ Watch out

Voices are solid, not spectacular — go elsewhere if you specifically need cloning or top realism.

📝 Deepgram (Nova-3) The most accurate cloud speech-to-text, batch or live

📝

Deepgram (Nova-3)

API key required deepgram.com →

by Deepgram · released Nova-3: 2025

The most accurate cloud speech-to-text, batch or live

Deepgram's Nova-3 model is widely regarded as the most accurate proprietary speech-to-text engine available: it has the lowest word-error-rate among commercial transcription APIs and handles both pre-recorded audio (batch) and live microphone streams (real-time) well. For a student this is the tool behind transcribing interviews, lab meeting recordings, or a live captioning feature in an app — anywhere accuracy on technical vocabulary and noisy audio matters more than cost. It's also cheap at scale: batch transcription runs about $0.0043 per minute of audio, and streaming about $0.0077 per minute, which is inexpensive compared to hiring a human transcriber but still an ongoing cost compared to running something locally. Like the other developer-facing tools in this chapter, using it means calling an API rather than clicking a button on a website.

Best for

Build an app

Also useful for

Draft docs

What it does well

Lowest word-error-rate of the major proprietary STT APIs
Handles both batch files and live streaming audio
Fast turnaround even on long recordings
Cheap per minute at scale

Where it falls short

Developer API only — needs code to use, no drag-and-drop web app
Audio leaves your machine and is processed in the cloud
Ongoing per-minute cost, unlike a one-time local setup

Try it in 5 minutes

Transcribe a recorded lab meeting

Upload a 5-minute recording of a lab meeting or lecture to the Deepgram API playground and compare the transcript against what you remember, especially technical terms.

Expected: A transcript with only a handful of errors, mostly on unusual jargon or overlapping speech.

Pricing

Batch transcription: ~$0.0043/min. Streaming transcription: ~$0.0077/min. Pay-as-you-go via API, with free trial credit.

⚠️ Watch out

It is an API for developers — there is no simple "upload a file, get text" website for casual use.

🎧 Whisper (open source) OpenAI's open-weight speech-to-text, running free on your laptop

🎧

Whisper (open source)

free github.com/openai/whisper →

by OpenAI (weights); whisper.cpp/faster-whisper (community runtimes) · released Whisper: 2022; whisper.cpp: 2022

OpenAI's open-weight speech-to-text, running free on your laptop

Whisper is OpenAI's speech-to-text model, and unlike ChatGPT the weights themselves were released openly (MIT license), so anyone can run the exact same model on their own computer for free, forever, with no internet connection and no data ever leaving the machine. Two community projects make this practical: whisper.cpp runs efficiently on ordinary CPUs and Apple Silicon (Metal), while faster-whisper uses a GPU to run about 4x faster than the original implementation. Both are the same underlying model as OpenAI's paid transcription API — you're not sacrificing much accuracy for the privilege of running it locally. This is the best open-source option for turning a recorded lecture, interview, or field recording into text, and a natural fit for a course that wants students comfortable with tools that cost nothing and don't require sending data to a company.

Best for

Open source / self-host
Build an app

Also useful for

Draft docs

What it does well

Completely free and runs offline, no data leaves your machine
Same model weights as OpenAI's paid API — comparable accuracy
whisper.cpp runs on ordinary laptop CPUs (and Apple Metal)
faster-whisper is ~4x faster if you have a GPU

Where it falls short

Command-line setup, not a polished point-and-click app
Local transcription is slower than a cloud API unless you have a decent GPU
No built-in speaker labeling or live-streaming mode out of the box

Try it in 5 minutes

Transcribe a lecture recording locally

Install whisper.cpp (or use a wrapper app for it), then run it on a recorded lecture or podcast episode and check the transcript against a section you remember well.

Expected: A working transcript produced entirely offline, with a setup step (installing the tool) before the first result.

Pricing

$0 — runs entirely on your own hardware once downloaded. No API key, no subscription, no per-minute cost.

⚠️ Watch out

You need to install and run software yourself — there is no polished app, just command-line tools.

🗣️ Kokoro 82M (open source) Tiny open TTS model that sounds far bigger than its 82M params

🗣️

Kokoro 82M (open source)

free huggingface.co/hexgrad/Kokoro-82M →

by hexgrad (community) · released 2025

Tiny open TTS model that sounds far bigger than its 82M params

Kokoro is a genuinely surprising result: an 82-million-parameter, Apache-2.0-licensed text-to-speech model that produces voice quality competitive with systems many times its size, while running comfortably on a CPU or a modest 2-4GB of GPU memory. That makes it realistic to run on a student laptop, embed in a small app, or self-host without renting a server. Its even lighter sibling, Piper, trades a little quality for extreme efficiency — it's designed to run on hardware as small as a Raspberry Pi, the natural choice if you're building something like an offline audio kiosk or an accessibility device. Worth knowing: some other open TTS models with excellent voice-cloning (XTTS-v2, F5-TTS) are released under non-commercial licenses, so they're fine for a course project but not for anything you'd sell or deploy commercially — Kokoro's Apache-2.0 license doesn't have that restriction.

Best for

Open source / self-host
Build an app

Also useful for

Draft docs

What it does well

Apache-2.0 license: free for commercial and non-commercial use alike
Only 82M parameters — runs on CPU or a small GPU
Voice quality punches well above its size
Piper sibling runs even lighter, down to Raspberry Pi hardware

Where it falls short

No voice cloning built in — you pick from provided voices, not your own
Fewer languages/voices than the big cloud providers
Rival OSS models with cloning (XTTS-v2, F5-TTS) are non-commercial-licensed — easy to reach for the wrong one

Try it in 5 minutes

Generate narration on your own laptop

Run Kokoro locally (via a Python script or an existing wrapper UI) on a paragraph of text and listen to the output with no internet connection active.

Expected: A clear, natural-sounding audio clip generated in a few seconds on ordinary laptop hardware, entirely offline.

Pricing

$0 — Apache-2.0 licensed, runs locally on CPU or 2-4GB VRAM. No API key or subscription required.

⚠️ Watch out

Non-commercial rivals like XTTS-v2 and F5-TTS clone voices better but cannot legally be used commercially — check licenses first.

🎧 OpenAI Realtime API Managed speech-to-speech API — talk to gpt-realtime directly

🎧

OpenAI Realtime API

paid platform.openai.com/docs/guides/realtime →

by OpenAI · released 2024 (GA 2025), gpt-realtime model 2025

Managed speech-to-speech API — talk to gpt-realtime directly

OpenAI's Realtime API lets an app stream microphone audio straight to the gpt-realtime model and get spoken audio back — no separate transcription or text-to-speech step to wire up yourself. Round-trip latency is typically 150-400ms, so it feels like an actual conversation rather than a walkie-talkie exchange. Under the hood it behaves like a single API even though speech understanding and generation are tightly coupled to the model, which is why tone, interruptions, and back-channel sounds ('mm-hm') come through more naturally than a bolted-together pipeline. It supports function calling mid-conversation, so the agent can look things up or trigger actions while still talking. This is the option to reach for when you want the most polished, least-assembly-required voice agent and don't mind paying a premium and staying inside OpenAI's ecosystem.

Best for

Talk through ideas
Automate workflow

Also useful for

Build an app

What it does well

Lowest-effort path to a natural-sounding voice agent
150-400ms latency — feels conversational
Native interruption handling (barge-in)
Function calling works mid-conversation

Where it falls short

Most expensive per-minute option in this lineup
Closed model — no self-hosting, no fine-tuning of the voice
Vendor lock-in to OpenAI billing and infrastructure

Try it in 5 minutes

Talk to the Realtime playground

Open the OpenAI Realtime console demo, pick a voice, and have a 2-minute spoken conversation asking it to explain photosynthesis, then interrupt it mid-sentence.

Expected: It answers in a natural voice within under half a second, and stops talking immediately when you interrupt.

Pricing

~$0.30/min headline rate; realistic all-in cost with prompt caching is closer to $0.25-0.35/min. Billed via OpenAI API usage, no separate voice add-on.

⚠️ Watch out

Per-minute cost adds up fast for long conversations — budget before you build.

🔵 Google Gemini Live Cheapest full-stack managed realtime voice, unified per-token billing

🔵

Google Gemini Live

paid ai.google.dev/gemini-api/docs/live →

by Google DeepMind · released 2024, expanded 2025

Cheapest full-stack managed realtime voice, unified per-token billing

Gemini Live is Google's answer to the same problem as OpenAI's Realtime API — stream audio in, get spoken audio back — but priced roughly 3-4x cheaper per minute, which matters a lot once you're running voice agents at any real volume. Latency sits around 100-200ms, on par with or better than the competition. Billing is unified per-token (audio, text, and tool calls all count against the same token meter) rather than a separate per-minute voice tariff, which makes cost easier to reason about if you're already used to Gemini's text pricing. It integrates with the rest of the Gemini API, so tool use, grounding, and multimodal input (you can show it a video frame while talking) come for free. For a student prototyping a cheap always-on voice tutor or lab-notebook assistant, this is usually the best cost-to-quality ratio among the managed options.

Best for

Talk through ideas
Automate workflow

Also useful for

Build an app

What it does well

Cheapest managed speech-to-speech option by a wide margin
100-200ms latency
Unified token billing — one meter, not separate voice + LLM charges
Multimodal: can see and hear at the same time

Where it falls short

Voice quality slightly behind ElevenLabs in blind listening tests
Closed model, Google ecosystem lock-in
Rate limits tighter than text-only Gemini API

Try it in 5 minutes

Try a Gemini Live demo

Use Google AI Studio's Live API stream demo and ask it to quiz you on cell biology terms out loud for 3 minutes.

Expected: Fast, natural back-and-forth quizzing with correct interruption handling, at a noticeably lower running cost than the OpenAI equivalent.

Pricing

~$0.07-0.12/min all-in at typical usage, billed per input/output token rather than a flat per-minute voice rate.

⚠️ Watch out

Cheapest doesn't mean free — token-based billing can still surprise you on long sessions.

🎙️ ElevenLabs Conversational AI Best-in-class voice quality via STT → LLM → ElevenLabs TTS pipeline

🎙️

ElevenLabs Conversational AI

free tier elevenlabs.io/conversational-ai →

by ElevenLabs · released 2024

Best-in-class voice quality via STT → LLM → ElevenLabs TTS pipeline

ElevenLabs built its reputation on the most natural-sounding text-to-speech on the market, and its Conversational AI product wraps that voice engine into a full pipeline: speech-to-text, your choice of LLM, then ElevenLabs' own TTS for the reply. Because it's cascaded rather than a single speech-native model, it's easier to swap the LLM brain in and out, and easier to debug when something goes wrong — you can inspect the transcript at each stage. The trade-off is you're stacking three costs (STT, LLM, TTS) and three hops of latency, though ElevenLabs has tuned the pipeline hard to keep it feeling responsive. This is the pick when voice quality — the actual timbre, emotion, and naturalness of what comes out of the speaker — matters more to your use case than shaving the last 100ms of latency.

Best for

Talk through ideas

Also useful for

Automate workflow
Build an app

What it does well

Best raw voice quality and emotional range in this lineup
Cascaded design = easy to swap LLM or debug a specific stage
Free tier (15 min/mo) to prototype without paying
Large voice library and voice cloning available

Where it falls short

Cascaded = higher latency than speech-native models
Three cost centers stacked (STT + LLM + TTS)
Burst pricing (2x) can bite under heavy concurrent load

Try it in 5 minutes

Build a 5-minute conversational agent

In the ElevenLabs Conversational AI dashboard, create an agent with a default voice and system prompt "You are a friendly biology tutor," then talk to it about mitosis.

Expected: A voice that sounds close to a real human, with natural pauses — and you can see the STT transcript update live as you speak.

Pricing

$0.08/min base rate, bursts to $0.16/min under load, plus LLM cost passed through at your provider's rate. Free tier: 15 minutes/month.

⚠️ Watch out

Cascaded architecture (three stages chained) — you're also paying LLM cost on top of the voice fee.

☎️ Retell AI Transparent bring-your-own-API voice agent platform, popular for phone agents

☎️

Retell AI

API key required www.retellai.com →

by Retell AI · released 2023

Transparent bring-your-own-API voice agent platform, popular for phone agents

Retell AI doesn't try to be the model — it's the orchestration layer that wires your chosen speech-to-text, LLM, and text-to-speech providers together into a phone-ready voice agent, handling call routing, telephony (SIP/Twilio), conversation state, and interruption logic for you. You bring your own API keys for each component, so you keep full control and full visibility over exactly what each piece costs and how it behaves — nothing is hidden inside a black box. This transparency is why it's become popular specifically for phone-based agents (appointment booking, support lines, outbound calling), where reliability and auditability matter more than squeezing out the last few milliseconds. Vapi is a close competitor with a similar bring-your-own-stack philosophy, if you want to compare. Retell is the right choice when you want to control your stack precisely and need real telephone integration, not just a browser demo.

Best for

Automate workflow

Also useful for

Build an app
Talk through ideas

What it does well

Full transparency — you see and control every component's cost
Built for real telephony (inbound/outbound phone calls via SIP/Twilio)
Mix and match best-of-breed STT/LLM/TTS per use case
Mature tooling for call analytics and conversation flows

Where it falls short

More assembly required than a fully managed option
Total cost = Retell fee + 3 separate provider bills, harder to estimate upfront
No included voice/LLM — nothing works until you plug in your own keys

Try it in 5 minutes

Set up a phone agent skeleton

In Retell's dashboard, create an agent using a free-tier STT, LLM, and TTS provider, then place a test call and ask it to explain what a ribosome does.

Expected: You can see each stage (transcript, LLM response, generated audio) separately in the call log, and swapping the TTS provider takes one dropdown change.

Pricing

~$0.07/min for Retell's own orchestration layer, on top of which you pay your own STT/LLM/TTS provider costs directly.

⚠️ Watch out

You still need your own STT, LLM, and TTS API keys — Retell is the orchestration layer, not the voice itself.

🔧 Pipecat Open-source, pluggable voice pipeline: VAD → STT → LLM → TTS

🔧

Pipecat

free www.pipecat.ai →

by Daily.co (open source community project) · released v1.0 in April 2026

Open-source, pluggable voice pipeline: VAD → STT → LLM → TTS

Pipecat is an MIT-licensed open-source framework for building the same kind of cascaded voice pipeline ElevenLabs or Retell offer commercially — voice activity detection, speech-to-text, an LLM, and text-to-speech — except you own every piece and pay nothing beyond the component costs (or $0 if you run local models). It reached v1.0 in April 2026 and ships with 60+ integrations, so you can plug in whichever STT, LLM, and TTS you already trust, including fully local ones. It's transport-agnostic, meaning it can carry audio over WebRTC via Daily or LiveKit, or other transports, so it isn't tied to one hosting vendor. LiveKit Agents (Apache-2.0) is the closest sibling project — also open source, also WebRTC-based, worth comparing if you're choosing a foundation. Pipecat is the pick when you want maximum control, zero vendor lock-in, and are comfortable doing the integration work yourself.

Best for

Open source / self-host
Build an app

Also useful for

Automate workflow

What it does well

Fully open source (MIT) — no vendor lock-in, inspect everything
60+ integrations: swap any stage independently
Transport-agnostic (Daily, LiveKit, others)
Can run entirely free with local models

Where it falls short

Requires real setup and hosting — not a five-minute demo
You're responsible for reliability, scaling, and monitoring
Quality depends entirely on which components you choose

Try it in 5 minutes

Run the Pipecat quickstart

Clone the Pipecat quickstart repo, wire in a free STT and TTS provider plus a local or API LLM, and run the sample voice bot locally.

Expected: After following the setup steps (expect 20-30 minutes, not 2), you get a working voice loop running on your own machine with full visibility into each pipeline stage.

Pricing

$0 for the framework itself (MIT license). You pay only for whichever STT/LLM/TTS components and hosting you choose — can be $0 if fully local.

⚠️ Watch out

You assemble and self-host the pipeline — expect real setup time before your first working demo.

🇫🇷 Moshi (Kyutai) Open-source speech-native model — full-duplex, sub-200ms, EU-hosted

🇫🇷

Moshi (Kyutai)

free kyutai.org/moshi →

by Kyutai (French AI lab) · released 2024, ongoing updates

Open-source speech-native model — full-duplex, sub-200ms, EU-hosted

Moshi, from French lab Kyutai, is a genuinely different architecture from everything else on this page: instead of chaining separate speech-to-text, language model, and text-to-speech stages, it's one model trained end-to-end on audio that listens and speaks simultaneously (full-duplex), the way humans actually converse — it can be interrupted mid-word and react instantly because there's no hand-off between separate components. That collapses latency to under 200ms and makes overlapping speech and natural back-channeling ('yeah', 'right') possible in a way cascaded pipelines struggle to fake. It's released under CC-BY-4.0 and can be self-hosted entirely in the EU, which matters if data residency or privacy is a concern. Ultravox is the other notable open speech-native model, taking a slightly different approach (audio feeds directly into an LLM, skipping a separate ASR step). Both require a decent GPU to run at real-time speed — this is the frontier-research end of open source, not a weekend project.

Best for

Open source / self-host
Talk through ideas

Also useful for

Build an app

What it does well

True speech-native, full-duplex — sub-200ms, interruptible mid-word
Open weights (CC-BY-4.0), can be fully self-hosted in the EU
No separate ASR/TTS hop — one model does it all
Genuinely novel architecture worth understanding, not just using

Where it falls short

Requires a capable GPU for real-time performance
Rougher edges than commercial voices — less polished than ElevenLabs
Smaller community and fewer integrations than Pipecat

Try it in 5 minutes

Run the Moshi demo

Try Kyutai's hosted Moshi demo (or run the open checkpoint locally if you have a GPU) and deliberately talk over it mid-sentence.

Expected: It reacts to the interruption almost instantly and can genuinely overlap speech with you — a noticeably different feel from a cascaded pipeline politely waiting its turn.

Pricing

$0 — CC-BY-4.0 license, only cost is your own GPU/hosting hardware.

⚠️ Watch out

Needs a real GPU to run at usable speed — this is a research-grade model, not a plug-and-play service.

🧑‍💼 HeyGen Market-leading talking-head avatars — pre-recorded and realtime

🧑‍💼

HeyGen

paid www.heygen.com →

by HeyGen (Hong Kong / US) · released 2020

Market-leading talking-head avatars — pre-recorded and realtime

HeyGen is the tool most people mean when they say "AI avatar video." Upload a photo (or use a stock avatar), type or paste a script, pick a voice, and it renders a lip-synced talking-head video in minutes — no camera, no studio. Its "Avatar IV" model produces natural mouth movement and expression from a single still photo. Beyond pre-recorded clips, HeyGen also ships an Interactive Avatar / Streaming Avatar API: you can wire an avatar to an LLM and have it hold a live, low-latency video conversation — useful for kiosks, virtual presenters, or a "talking chatbot" front end. The free tier is enough to try it; anything resembling regular use (a course, a product demo library, a support bot) needs a paid plan. Good starting point for students because the pre-recorded workflow requires zero technical setup — it is the closest thing to "PowerPoint with a face."

Best for

Draft docs

Also useful for

Build an app

What it does well

Best-in-class lip sync and expression from a single photo
Both pre-recorded AND realtime interactive in one product
40+ languages with automatic dubbing
Fastest path from script to video for non-technical users

Where it falls short

Realtime avatar minutes are expensive at scale
Free tier is watermarked and very limited
Photorealistic likeness raises consent questions if it is not your own face

Try it in 5 minutes

Make a 20-second talking-head clip

Upload a headshot photo, paste a 3-sentence script introducing yourself, pick a voice, generate.

Expected: A lip-synced video of "you" (or the stock avatar) speaking your script, ready in under 5 minutes.

Pricing

Creator $29/mo (~10 min/mo Avatar IV); Pro $99/mo; Business $149/mo + $20/seat; API $0.05/sec; Interactive Avatar ≈ $0.10-0.20/min (~$6-12/hr).

⚠️ Watch out

Realtime "Interactive Avatar" is billed separately and per-minute — costs add up fast in a live session.

🎬 Synthesia Enterprise training-video avatars with PowerPoint-to-video and SCORM export

🎬

Synthesia

paid www.synthesia.io →

by Synthesia (UK) · released 2019

Enterprise training-video avatars with PowerPoint-to-video and SCORM export

Synthesia targets corporate training and internal communications rather than individual creators. Its signature feature is turning a PowerPoint deck or a script into a full training video with a presenter avatar, in 140+ languages, with one click to export as SCORM for an LMS. It is the tool an HR or L&D team reaches for to produce "here is our new policy" videos at scale without hiring actors or camera crews. Compared to HeyGen, Synthesia leans further into templates, brand kits, and team collaboration features (approval workflows, shared avatar libraries) rather than experimentation. It does not offer a realtime interactive avatar product — everything is pre-rendered. For students, the useful lesson here is less "try this yourself" and more "this is what happens when avatar video gets adopted inside a company": localization at scale, compliance-friendly training content, and video produced by people with zero video-editing skill.

Best for

Draft docs

What it does well

Script or slide deck straight to finished video
140+ languages with built-in translation
SCORM export for LMS / corporate training platforms
Team workflows: brand kits, shared avatars, approvals

Where it falls short

No realtime/interactive avatar option
Pricing and UX are built for teams, not casual individual use
Avatars are polished but can feel corporate/generic

Try it in 5 minutes

Turn a slide into a narrated video

Paste 3 bullet points as a script and generate a 30-second presenter video from a stock avatar.

Expected: A clean, corporate-style talking-head video narrating your bullets, exportable as MP4.

Pricing

Starter $29/mo (~10 min of video); Business tiers scale per minute of output, custom enterprise pricing for larger volumes.

⚠️ Watch out

Built for corporate L&D workflows — overkill if you just want one quick clip.

🗣️ D-ID Photo-to-talking-head plus sub-0.5s realtime LLM-connected agents

🗣️

D-ID

paid www.d-id.com →

by D-ID (Israel) · released 2017

Photo-to-talking-head plus sub-0.5s realtime LLM-connected agents

D-ID made its name animating a single photo into a talking head — feed it any portrait and a script or audio clip, and it lip-syncs the face to the speech. That core "Creative Reality Studio" is the pre-recorded side. The more interesting recent development is D-ID Agents (V4 Visual Agents): a realtime avatar wired directly to an LLM, with sub-0.5-second response latency and 4K video, aimed at customer-service and website-greeter use cases. It is a direct competitor to Tavus and HeyGen's interactive product in that space. D-ID's entry price is the lowest of the paid proprietary tools, making it a reasonable second stop for students after HeyGen to compare photo-animation quality and see a different take on the "embed a talking avatar on your website" pattern.

Best for

Draft docs

Also useful for

Build an app

What it does well

Very low entry price for basic photo-to-video
Realtime Agents hit sub-0.5s latency, genuinely conversational
4K output on the agent product
Simple API for developers who want to embed a talking face

Where it falls short

Free/cheap tier output is watermarked and lower resolution
Agent pricing can surprise you at conversation scale
Photo-animation quality is good but not as expressive as HeyGen Avatar IV

Try it in 5 minutes

Animate a still photo

Upload any front-facing photo plus a short audio clip or typed script.

Expected: A talking-head video where the photo's mouth and head move in sync with the speech.

Pricing

Entry plan ≈ $4.70/mo billed annually; Visual Agents priced on usage (per conversation/minute), scales with volume.

⚠️ Watch out

Cheapest entry price of the proprietary tools, but usage-based agent pricing scales with conversation volume.

👤 Tavus (CVI) The realtime interactive avatar specialist — low-latency conversational video

👤

Tavus (CVI)

free tier www.tavus.io →

by Tavus (US) · released 2021

The realtime interactive avatar specialist — low-latency conversational video

Tavus built its Conversational Video Interface (CVI) specifically for realtime, back-and-forth video conversation with an AI avatar — this is the category leader for "talk to a video AI face and have it respond naturally," not for generating pre-recorded clips. You give it a persona and connect it to an LLM; it handles the low-latency video generation so the avatar listens, thinks, and replies on camera with normal conversational timing. Growth-tier plans support many concurrent sessions, which matters if you are building something like a virtual tutor or interviewer that many students talk to at once. Free tier (25 min) is enough to feel the difference between this and a pre-recorded HeyGen clip. Worth mentioning: Simli is a much cheaper realtime alternative (~$0.009/min) for teams that want the interactive pattern without Tavus's price tag, at some cost in polish.

Best for

Build an app

Also useful for

Talk through ideas

What it does well

Purpose-built for realtime, low-latency conversation (the actual hard problem)
Free tier lets you test genuine interactivity before paying
Concurrent-session support for many simultaneous users
Developer-first API designed to sit behind any LLM

Where it falls short

Not designed for producing a library of pre-recorded videos
Growth tier pricing is steep for student/hobbyist budgets
Realtime avatar conversation can still feel slightly uncanny under fast back-and-forth

Try it in 5 minutes

Have a live conversation with an avatar

Use the free tier to start a CVI session and ask the avatar three follow-up questions in a row.

Expected: The avatar responds on video in near-real-time, holding a natural back-and-forth — not a static playback.

Pricing

Free tier: 25 min; Starter $59/mo; Growth $395/mo (15 concurrent sessions). Cheaper realtime alt: Simli ≈ $0.009/min.

⚠️ Watch out

Optimized for live conversation, not for producing a polished pre-recorded video library.

🦾 SadTalker Open-source single-photo talking head — free, self-hosted, needs a GPU

🦾

SadTalker

free github.com/OpenTalker/SadTalker →

by OpenTalker (research / open source community) · released 2023

Open-source single-photo talking head — free, self-hosted, needs a GPU

SadTalker is the open-source answer to "animate one photo into a talking head": feed it a still image and an audio clip, and it generates 3D-aware head motion (not just mouth flapping) synced to the audio, using a research pipeline published alongside an academic paper. It runs entirely on your own machine or a rented GPU — no subscription, no per-minute fee, no vendor. The tradeoff is exactly what you would expect from open source: you need to set up a Python environment, download model weights, and have a GPU with enough VRAM, or rent one (a cloud GPU for an hour or two is cheap but is a real step most SaaS users skip). Quality is good but not quite at HeyGen or D-ID polish level. For students, this is the "how does the free version actually work" lesson — and a stepping stone to LivePortrait (Kuaishou), a newer open-source sibling that pushes closer to realtime with noticeably better motion quality, if you want the more advanced option.

Best for

Open source / self-host

What it does well

Completely free, no vendor lock-in or per-minute fees
3D-aware head motion, not just lip movement
Runs fully offline once set up
Active open-source ecosystem (LivePortrait as a higher-quality sibling)

Where it falls short

Requires a GPU — slow or impractical on CPU-only machines
Setup requires comfort with Python/command line
Output quality trails the top paid tools
No production-ready realtime version yet

Try it in 5 minutes

Run it on your own headshot

Clone the repo, feed it a headshot photo and a short .wav audio clip, run inference.

Expected: A talking-head video generated locally — free, but you will feel the GPU/setup requirement firsthand.

Pricing

Free — self-hosted; costs only what your own GPU (local or rented) costs to run.

⚠️ Watch out

No hosted service — you run it yourself, and it wants a decent GPU or it will be painfully slow.

💋 LatentSync / Wav2Lip Open-source lip-sync for dubbing existing video into another language

💋

LatentSync / Wav2Lip

free github.com/bytedance/LatentSync →

by LatentSync: ByteDance; Wav2Lip: IIIT Hyderabad research team (open source) · released Wav2Lip 2020; LatentSync 2024

Open-source lip-sync for dubbing existing video into another language

This is the open-source pair for a different, very practical job: taking a video you already have and re-syncing the speaker's mouth to a new audio track — most commonly to dub a talk or a lecture into another language convincingly. LatentSync (ByteDance) is the newer, latent-diffusion-based approach and produces noticeably better mouth and teeth detail than older methods. Wav2Lip is the classic GAN-based tool from 2020 that is still widely used because it is fast, lightweight, and works across languages without retraining. Both are free and self-hosted, with the same tradeoffs as SadTalker: you need a GPU and are comfortable running a Python project rather than clicking a web UI. For the course, the practical framing is: if you already have a real video of a real lecture and just want it in Spanish or German with matching lip movement, this is the free open-source path — HeyGen and Synthesia can do this too, but as a paid, hosted feature.

Best for

Open source / self-host

What it does well

Free and self-hosted, no per-minute dubbing fees
LatentSync gives noticeably sharper mouth/teeth detail than older methods
Wav2Lip is fast, lightweight, and battle-tested across many languages
Good for dubbing real footage, not just synthetic avatars

Where it falls short

Requires a GPU and Python setup, no polished web UI
Re-syncs an existing video — cannot create a new talking-head from scratch
Wav2Lip output looks dated next to newer diffusion-based methods

Try it in 5 minutes

Dub a short clip into a new audio track

Take a 10-second video of someone talking and a new audio recording of different words; run LatentSync or Wav2Lip to re-sync the mouth.

Expected: The original video, but the mouth now matches the new audio track — a basic dubbing pipeline, for free.

Pricing

Free — self-hosted; costs only your own GPU (local or rented).

⚠️ Watch out

These re-sync mouths on existing footage — they do not generate a person or a video from scratch.

🎥 Sora 2 Cinema-grade text-to-video with character consistency

🎥

Sora 2

paid openai.com/sora →

by OpenAI · released 2025 (Sora 2, 2026 updates)

Cinema-grade text-to-video with character consistency

Sora 2 is OpenAI's flagship text-to-video model: describe a scene in a sentence and it renders up to 60 seconds of footage at 720p (1080p on the Pro tier), holding a character's face and clothing consistent across cuts — the single hardest problem in AI video. It's built into ChatGPT Plus ($20/month) and Pro ($100/month) as a simple prompt box, and exposed as an API for developers who want to generate video programmatically. For a student this looks like: paste a description of a cell dividing or a protein folding, get back a short clip you can drop into a slide deck. The catch is that Sora doesn't generate audio — you still need a separate voice or music tool — and OpenAI has already signaled the API layer is not a long-term bet, so treat this as a snapshot of the state of the art, not a permanent platform choice. Cinematic quality is the headline feature: camera moves, lighting, and physics look convincingly like real footage rather than a slideshow of frames.

Best for

Draft docs

What it does well

Best-in-class character and scene consistency across a 60-second clip
1080p output on the Pro tier — sharp enough for a real presentation
Simple prompt-box UX inside ChatGPT — no separate account if you already pay for Plus
Strong physical realism: lighting, camera motion, and object interaction look natural

Where it falls short

No built-in audio — every clip is silent, you add sound separately
API access is flagged for sunsetting (~Sept 2026) — don't build a workflow that depends on it long-term
Paid only; no free tier to experiment with beyond ChatGPT Plus

Try it in 5 minutes

Generate a 5-second clip from text

In ChatGPT (Plus/Pro), open Sora and type: "A close-up of a plant cell dividing, mitosis in progress, soft lab lighting, slow zoom in." Generate a 5-second clip.

Expected: A short, silent video clip you can watch and download — realistic lighting and camera motion, ready to drop into a slide.

Pricing

Included in ChatGPT Plus ($20/mo) and Pro ($100/mo) with usage limits; API access billed per second — $0.10/sec on the base model, $0.30-0.50/sec on Sora 2 Pro.

⚠️ Watch out

OpenAI has flagged the Sora API for sunsetting around September 2026 — learn the category and technique, not just this one vendor.

🎞️ Google Veo 3 Fast, budget text-to-video with native audio

🎞️

Google Veo 3

paid deepmind.google/technologies/veo/ →

by Google DeepMind · released 2025 (Veo 3), 2026 (Veo 3.1)

Fast, budget text-to-video with native audio

Veo 3 is Google's text-to-video model, built for speed and prompt-adherence rather than maximum length: you type a description and get a roughly 6-8 second clip at close to 1080p in well under a minute. Its standout feature versus Sora and Runway is native audio — Veo 3 generates ambient sound, sound effects, and even dialogue synced to the video in the same pass, so you don't need a separate audio tool for a simple clip. It's accessible through Google AI Pro (~$29/month, bundled with Gemini) or pay-per-clip via the lighter Veo 3.1 Light tier (~$0.05/clip), which makes it one of the cheaper ways to experiment. For a student, Veo is the pragmatic choice for a single illustrative shot — a heartbeat visualization, a molecule rotating — where you want decent quality fast and don't need a long cinematic sequence. It follows prompts unusually literally, which is good for precision but means vague prompts produce vague results.

Best for

Draft docs

What it does well

Native audio generation — sound effects and dialogue synced to video, no separate tool needed
Fast generation and strong prompt-adherence — what you type is close to what you get
Cheapest per-clip cost among the proprietary frontier models
Accessible via a Google AI Pro subscription you may already have for Gemini

Where it falls short

Short clips only — about 6-8 seconds, not suited to a multi-shot sequence
Less cinematic camera control than Runway or Sora for filmmaking-style work
Paid only, though the Light tier is inexpensive to trial

Try it in 5 minutes

Generate a clip with sound in one pass

In Gemini / Google AI Pro, open Veo and type: "A heart beating in a translucent blue glow, ambient heartbeat sound, slow rotation." Generate the clip.

Expected: A ~6-8 second clip with synced audio — no separate sound step required.

Pricing

Bundled in Google AI Pro (~$29/mo) via Gemini; Veo 3.1 Light runs about $0.05 per clip on pay-as-you-go pricing.

⚠️ Watch out

Clips top out around 6-8 seconds — fine for a b-roll shot, not a scene.

🎬 Runway Gen-4 Pro filmmaking tools: motion control, multi-shot consistency

🎬

Runway Gen-4

paid runwayml.com →

by Runway (Cristóbal Valenzuela, NYC) · released 2025 (Gen-4)

Pro filmmaking tools: motion control, multi-shot consistency

Runway is the tool professional filmmakers and VFX artists reach for: Gen-4 generates roughly 12 seconds at up to 1080p, but its real strength is control — you can direct camera motion explicitly, keep a character or object consistent across multiple separate shots, and combine generated footage with real footage in the same timeline. It grew out of Runway's earlier green-screen and motion-tracking AI tools, so the interface assumes you're editing a sequence, not just prompting a single clip. Pricing starts at $12/month for the Standard plan (625 credits), with higher tiers scaling credits for heavier use. For a student this is overkill for a single illustrative shot but the right choice if you're building a short explainer video with several connected scenes — say, a multi-step process like a virus entering a cell, replicating, and being cleared by the immune system, all as one consistent visual sequence.

Best for

Draft docs

What it does well

Explicit motion/camera control — direct pans, zooms, and tracking shots
Multi-shot character and scene consistency across a sequence, not just one clip
Mature editing toolset built for combining generated and real footage
Lower cost of entry ($12/mo) than a full production suite

Where it falls short

Steeper learning curve — the tool assumes a filmmaking workflow, not a one-shot prompt
Credits get consumed fast on multi-shot or iterative projects
No native audio — sound is a separate step

Try it in 5 minutes

Generate one directed shot

In Runway Gen-4 (free trial credits), type: "A single algae cell drifting in water, camera slowly pans left to right, soft green backlighting." Generate and inspect the camera motion.

Expected: A ~5-10 second clip where the camera actually follows your directed pan — more controllable than a plain text-to-video prompt.

Pricing

Standard plan $12/month for 625 credits; Pro and higher tiers scale credit allowances for heavier production use.

⚠️ Watch out

Built for people doing real video production — more control surface than a one-off clip needs.

🐉 Kling Physics-accurate video at the lowest proprietary cost

🐉

Kling

paid klingai.com →

by Kuaishou (China) · released 2024, ongoing updates through 2026

Physics-accurate video at the lowest proprietary cost

Kling, from China's Kuaishou, has built a reputation for physically plausible motion — water splashing, cloth folding, objects colliding — at a noticeably lower price than the US frontier models. It generates roughly 10 seconds of 1080p video, and at around $0.84 per 10-second clip pay-as-you-go (or about $12/month bundled), it's one of the cheapest ways to get proprietary-grade quality. For a student on a budget, Kling is a strong default: it handles fluid dynamics and physical interactions — think blood flow, cell membrane deformation, particles moving through a medium — more convincingly than models optimized purely for cinematic look. It doesn't have the brand recognition or English-language tutorial ecosystem of Sora or Runway yet, so expect a bit more trial-and-error to find prompt patterns that work well.

Best for

Draft docs

What it does well

Notably strong physical accuracy — fluids, cloth, and collisions look believable
Cheapest cost-per-clip among the proprietary frontier tools
1080p output at a full 10-second clip length
Good prompt-adherence for physical/mechanical scenes

Where it falls short

Smaller English-language community and tutorial base than Sora or Runway
Interface and support are less polished for non-Chinese-speaking users
No native audio generation

Try it in 5 minutes

Test physical realism

In Kling, type: "A drop of dye spreading through a glass of water, slow motion, macro lens." Generate a 10-second clip.

Expected: A clip where the dye diffuses and swirls with believable fluid physics — a good test of what Kling does best.

Pricing

Approximately $0.84 per 10-second clip pay-as-you-go, or roughly $12/month for a bundled credit plan.

⚠️ Watch out

Newer entrant with a smaller ecosystem than Sora/Runway — fewer tutorials, less third-party tooling.

🌀 Wan 2.2 Open-source (Apache-2.0), the most versatile self-hosted video model

🌀

Wan 2.2

free open source github.com/Wan-Video/Wan2.2 →

by Alibaba Tongyi Lab · released 2025 (Wan 2.2)

Open-source (Apache-2.0), the most versatile self-hosted video model

Wan 2.2 is Alibaba's Tongyi Lab open-source video model, released under a permissive Apache-2.0 license — you can download the weights, run them on your own hardware, and use the output commercially with no per-clip fee. It uses a mixture-of-experts (MoE) architecture and is the most versatile open model available: it handles both text-to-video and image-to-video (animate a still photo), producing 720p-1080p clips around 5 seconds long. This is the model to reach for if you want to understand how video generation actually works under the hood, or if you're generating enough clips that per-clip API costs would add up. The real cost is hardware: you need a GPU with at least 24GB of VRAM to run it comfortably, which usually means a rented cloud GPU (see the Inference providers chapter) rather than a laptop. Setup typically goes through ComfyUI, a node-based interface for running these models.

Best for

Open source / self-host

Also useful for

Draft docs

What it does well

Fully open (Apache-2.0) — free to run, modify, and use commercially, no vendor lock-in
Handles both text-to-video and image-to-video in one model — most versatile OSS option
Active community and ComfyUI integration for setup and workflows
No per-clip cost once you have GPU access — good for high-volume generation

Where it falls short

Requires a 24GB+ VRAM GPU — not runnable on a typical laptop, needs a rented cloud GPU
Clips are shorter (~5s) than the top proprietary models
Setup (ComfyUI, model downloads, dependencies) has a real learning curve versus a hosted app

Try it in 5 minutes

See it running (no local GPU needed)

Rent a GPU instance (see Inference providers chapter), install ComfyUI with the Wan 2.2 workflow, and generate a 5-second clip from a text prompt of your choice.

Expected: A self-hosted clip generated entirely on infrastructure you controlled — the trade-off is setup time versus zero per-clip cost.

Pricing

$0 for the model itself (Apache-2.0, self-hosted); real cost is GPU compute — a rented 24GB+ GPU runs roughly $0.50-2/hour depending on provider.

⚠️ Watch out

Free to run, but needs a serious GPU (24GB+ VRAM) — this is a compute cost, not a money cost.

🎇 HunyuanVideo Large open text-to-video model, needs a serious GPU

🎇

HunyuanVideo

free open source github.com/Tencent/HunyuanVideo →

by Tencent · released 2024, updated through 2026

Large open text-to-video model, needs a serious GPU

HunyuanVideo is Tencent's large open-source text-to-video model, built as a full-attention transformer rather than the more compute-efficient architectures some competitors use. That architectural choice buys quality — 1280x720 output with strong visual fidelity for around 5-second clips — at the cost of hardware demands: you need 60GB or more of VRAM to run it, which in practice means a multi-GPU rented setup, not a single consumer card. It's free to download and use (no per-clip fee), and like Wan 2.2 it's typically run through ComfyUI. For a student, HunyuanVideo is worth knowing about as the 'high-end open' option — useful if you're already renting serious GPU capacity for other work (training a model, running a large LLM locally) and want to add video generation to the same rig, but Wan 2.2 or LTX-Video are more practical starting points given the lower hardware bar.

Best for

Open source / self-host

What it does well

Strong visual quality from a full-attention transformer architecture
1280x720 resolution output
Fully open weights — free to use and modify
Active open-source ecosystem and ComfyUI support

Where it falls short

Needs 60GB+ VRAM — the steepest hardware requirement of the open models here
Clips are short (~5s), similar to other open models
Less versatile than Wan 2.2 (no strong image-to-video path)

Try it in 5 minutes

Understand the hardware bar before renting

Look up a HunyuanVideo ComfyUI workflow guide and check the VRAM requirement against a GPU rental provider's pricing page before generating anything.

Expected: A clear picture of what a HunyuanVideo run actually costs in rented GPU time — the real barrier for this model.

Pricing

$0 for the model (open weights, self-hosted); GPU rental for 60GB+ VRAM typically runs several dollars per hour.

⚠️ Watch out

Needs 60GB+ VRAM — this is a multi-GPU or heavy-rental setup, not a hobbyist card.

🎼 LTX-Video Fast open video — first OSS model with synced audio in one pass

🎼

LTX-Video

free open source github.com/Lightricks/LTX-Video →

by Lightricks · released 2024, audio+video sync added 2026

Fast open video — first OSS model with synced audio in one pass

LTX-Video, from Lightricks (the company behind the Facetune and video-editing apps), is built for speed and accessibility rather than maximum fidelity: it runs on more modest GPUs than Wan 2.2 or HunyuanVideo, making it the most approachable open video model to actually try yourself. Its headline feature, added in 2026, is being the first open-source model to generate synced audio and video together in a single pass — closing a gap that used to mean every open video clip was silent by default. It produces 768p-1080p clips up to about 24 seconds, longer than most competitors in either open or proprietary categories. For a student without access to a 60GB GPU, LTX-Video is often the realistic entry point into self-hosted video generation: lower hardware bar, longer clips, and — increasingly — sound baked in, all for $0 beyond the GPU rental.

Best for

Open source / self-host

Also useful for

Draft docs

What it does well

First open model with synced audio+video generation in one pass
Longer clips than most alternatives — up to ~24 seconds
Runs on more modest GPUs — lower barrier to entry than Wan 2.2 or HunyuanVideo
Fast generation speed, good for iterating on prompts

Where it falls short

Audio+video sync is a new (2026) feature — quality and reliability still vary by prompt
Visual fidelity generally trails HunyuanVideo's full-attention approach
Still requires GPU rental/ownership and ComfyUI-style setup

Try it in 5 minutes

Generate a clip with audio on modest hardware

Rent a mid-tier GPU, set up LTX-Video via ComfyUI, and generate a 10-second clip with a text prompt that includes an ambient sound description (e.g. "birds chirping in a forest").

Expected: A clip with both video and roughly-synced audio generated in one pass — on cheaper hardware than the other open models here.

Pricing

$0 for the model (open weights, self-hosted); runs on more modest/cheaper rented GPUs than Wan 2.2 or HunyuanVideo.

⚠️ Watch out

The first open model to do audio+video together (2026) — still a young feature, quality varies by prompt.

🖼️ GPT Image 2 OpenAI's flagship image model — sharp text, 2K output, ChatGPT-native

🖼️

GPT Image 2

paid openai.com/index/introducing-chatgpt-images-2-0/ →

by OpenAI · released April 2026 (gpt-image-2)

OpenAI's flagship image model — sharp text, 2K output, ChatGPT-native

GPT Image 2 (API model id gpt-image-2, product name "ChatGPT Images 2.0") is OpenAI's third-generation native image model, released 21 April 2026. It topped every major Image Arena leaderboard immediately on release, with its biggest jump in text-to-image prompt adherence — it reads and renders long, correctly-spelled in-image text (labels, multi-language captions, diagram callouts) far better than any prior OpenAI model. It also does "structured generation": ask for a labeled scientific diagram or a step-by-step infographic and it composes the layout, not just a picture. Access is either inside ChatGPT (Plus/Pro, usage-limited) with zero setup, or via the API/Codex for programmatic, higher-volume use. For a Heidelberg lab or founder, its sweet spot is any image that needs to carry real, legible text — a labeled cell-cycle diagram, an annotated architecture sketch, a slide figure — a category most diffusion-only models still get wrong.

Best for

Draft docs

What it does well

Reached #1 across Image Arena leaderboards on release, with the largest-ever recorded lead specifically on text-to-image prompt/text rendering
Renders long, accurate in-image text in multiple languages — labels, diagrams, multi-panel infographics
No separate account or tool needed if you already pay for ChatGPT Plus/Pro
Structured generation: ask for a diagram or infographic and get a composed layout, not just a picture

Where it falls short

API pricing is metered per token, not per image — costs compound quickly across iterative editing rounds
OpenAI's standard API terms carry no copyright/IP indemnification for commercial output
Closed model: no local/offline option, no fine-tuning, no visibility into training data

Try it in 5 minutes

Generate a labeled diagram from a plain-English prompt

In ChatGPT (Plus/Pro, or free with limits): "Generate a labeled diagram of the cell cycle stages (G1, S, G2, M) as a clean scientific figure, white background, sans-serif labels."

Expected: A publication-style figure with correctly spelled, correctly positioned labels — the strongest current model for images that must carry embedded text.

Pricing

API is token-based: $8.00/1M image input tokens, $30.00/1M output tokens, $5.00/1M text input tokens (roughly $0.03 per 1K-resolution image, up to ~$0.06 at 4K); included with usage limits inside the existing ChatGPT Plus/Pro subscription tiers (list prices are well-known public figures, not re-verified by direct fetch this session).

⚠️ Watch out

API access is billed per token, not per image, so iterative editing loops add up fast — and OpenAI's standard terms carry no copyright/IP indemnification for the output.

🎨 Midjourney V8.1 The aesthetic leader — subscription only, no free tier, no API

🎨

Midjourney V8.1

paid www.midjourney.com →

by Midjourney, Inc. (David Holz) · released V8.1 became the default model June 2026 (Midjourney itself launched 2022)

The aesthetic leader — subscription only, no free tier, no API

Midjourney remains the aesthetic leader of the field: its default output looks "art-directed" with the least prompt effort of any tool here. V8.1 became the default model on 11 June 2026 (superseding V7), rendering 4-5x faster and at roughly double the resolution/4x the detail of V7's HD mode, while keeping V7's style-reference and Omni Reference consistency features. There is no API and no pay-as-you-go option — every image is billed against a required monthly subscription measured in "fast GPU hours," not image count (a typical still costs about 1 GPU-minute; an HD video batch costs roughly 26). Access is via Discord or the midjourney.com web app. For a founder or lab building anything visual-first — a pitch deck hero image, a product mockup, a conference poster — it is usually the fastest way to something that looks finished without heavy prompt-engineering.

Best for

Draft docs

What it does well

The most distinctive, art-directed default aesthetic of any model here — needs the least prompt-engineering to look "finished"
V8.1 (June 2026) renders 4-5x faster than earlier versions and roughly doubled HD resolution vs V7
Omni Reference and style references (--sref) give repeatable character/brand consistency across a shoot
A huge, searchable community prompt archive — the fastest way to learn technique by example

Where it falls short

No API and no free trial — payment is required before the first image
Discord-first workflow (a web app exists, but Discord is still where most power users operate) is a real onboarding hurdle
Organizations over $1M annual revenue must be on the Pro or Mega plan for compliant commercial use

Try it in 5 minutes

Generate a lab-poster hero image

/imagine a glowing 3D molecule model floating above a lab bench, soft blue lighting, scientific poster style --ar 16:9

Expected: Four highly polished, art-directed variations you can upscale — no other tool on this list gives this level of default visual polish.

Pricing

Four subscription tiers, no pay-as-you-go: Basic $10/mo (3.3 fast GPU hrs, no Relax mode), Standard $30/mo (15 hrs, unlimited Relax), Pro $60/mo (30 hrs + private "Stealth" generations), Mega $120/mo (60 hrs); 20% off with annual billing. Unused fast-hours do not roll over.

⚠️ Watch out

No free trial and no API — you must pay before generating a single image, and companies over $1M annual revenue must be on the Pro or Mega plan for compliant commercial use.

🔓 FLUX.2 The open one — Apache-2.0 weights up to a frontier-grade paid API

🔓

FLUX.2

free tier open source bfl.ai →

by Black Forest Labs · released November 2025 (FLUX.2); klein variant added January 2026

The open one — Apache-2.0 weights up to a frontier-grade paid API

FLUX.2 (Black Forest Labs, the team behind the original Stable Diffusion) is a frontier image family released November 2025 and split into four tiers by openness: klein (added January 2026 — a 4B model under Apache 2.0, fully free and commercially unrestricted, runs on a single consumer GPU like an RTX 3090/4070), dev (a 32B open-weight model that needs a paid commercial license to use beyond research), and flex/pro/max (proprietary, API-only, the sharpest photorealism and multi-image editing). That spread — a genuinely free Apache-2.0 tier all the way up to an enterprise API — makes it the natural "which one is actually free" answer among the closed frontier vendors, and it runs natively in ComfyUI from day one. Its standout capability is multi-reference editing: feed it several reference images and one instruction, and it composes a single coherent edit.

Best for

Draft docs

Also useful for

Open source / self-host

What it does well

The only frontier-tier family with a truly free, Apache-2.0, commercially-unrestricted variant (klein 4B) runnable on a consumer GPU
Strong multi-reference image editing — combine several reference images into one coherent edit in a single instruction
Runs natively in ComfyUI on day one — the same node-graph workflow used for Stable Diffusion
Published, predictable per-megapixel API pricing instead of an opaque credit system

Where it falls short

The best-quality tiers (pro/flex/max) are API-only and proprietary — not self-hostable at any price
dev, the strongest self-hostable tier, needs a paid commercial license ($999/mo+) once you're not purely researching
Smaller fine-tune/LoRA ecosystem than Stable Diffusion at this point

Try it in 5 minutes

Run FLUX.2 klein locally for free

Install ComfyUI, load the FLUX.2-klein-4B workflow from Black Forest Labs' Hugging Face repo, and prompt: "a researcher's desk with a laptop showing a DNA double helix, natural window light."

Expected: A generation rendered entirely on your own GPU — no account, no per-image cost, no rate limit.

Pricing

klein 4B: free, Apache 2.0, self-hosted. Paid API, roughly per-megapixel: klein from ~$0.014/image, klein 9B ~$0.015, pro from ~$0.03, flex ~$0.05, max ~$0.07. Self-hosted commercial use of dev requires a separate license starting at $999/mo (100,000 images/mo included, $0.01/image after).

⚠️ Watch out

Only the smallest variant (klein 4B) is genuinely free and unrestricted — the stronger self-hostable tier (dev) needs a paid commercial license once you're past research use, and the very best quality (pro/flex/max) is API-only at any price.

🍌 Nano Banana Pro (Gemini 3 Pro Image) Google's multimodal reasoning image model — grounded, accurate text and data

🍌

Nano Banana Pro (Gemini 3 Pro Image)

paid ai.google.dev/gemini-api/docs/pricing →

by Google DeepMind · released November 2025 (preview); general availability June 2026

Google's multimodal reasoning image model — grounded, accurate text and data

"Nano Banana Pro" is Google's public nickname for Gemini 3 Pro Image, released as a preview in November 2025 and reaching general availability in June 2026. Unlike a single-purpose diffusion model, it runs on Google's Gemini 3 multimodal reasoning stack — so it can ground an image in real facts pulled via Search (an accurate chart of real data, a correctly structured molecule diagram), follow long multi-part instructions in one pass, and render dense, accurate text and data labels across languages. It is Google's direct replacement for the Imagen line, which is being retired: Imagen 4's endpoints shut down 17 August 2026, with Google recommending migration to this Gemini image family (the cheaper sibling is gemini-3.1-flash-image). Access is via the Gemini app, the Gemini API, or Vertex AI — useful if a lab already runs on Google Cloud/Workspace.

Best for

Draft docs

What it does well

Grounds images in real facts via Search-backed reasoning (an accurate chart of real data, a correctly-labeled real structure) rather than pure pattern-matching
Best-in-class dense, multi-language text rendering — full paragraphs and data labels, not just short captions
Follows long, multi-constraint prompts (compose several sub-scenes, keep one character consistent, add three specific labels) in a single generation
Direct API + Gemini app + Vertex AI access slots straight into an existing Google Cloud/Workspace setup

Where it falls short

Zero free tier — every generation is billed, even a first test image
Google is actively migrating its whole image line onto this family, so names/prices are likely to shift again within the year
Token-based pricing is harder to budget for upfront than a flat per-image rate

Try it in 5 minutes

Generate an accurate, labeled chart from real data

In the Gemini app (3 Pro): "Create a bar chart image comparing 2020-2025 global renewable energy capacity by source, with accurate axis labels and a legend, clean infographic style."

Expected: A genuinely data-grounded chart with correct layout and labels (verify the actual numbers before publishing) — a category pure diffusion models regularly get wrong.

Pricing

Token-based: $2.00/1M input tokens, $12.00/1M text-output tokens, $120.00/1M image-output tokens — roughly $0.134 per 1K/2K image and $0.24 per 4K image; the batch API roughly halves both rates.

⚠️ Watch out

No free tier at all, and Google is mid-migration off its older Imagen line onto this Gemini-native family (Imagen 4 endpoints shut down 17 Aug 2026) — expect further naming/pricing churn this year.

🕸️ Stable Diffusion 3.5 + ComfyUI Fully local, fully free — the open node-graph workflow

🕸️

Stable Diffusion 3.5 + ComfyUI

free open source huggingface.co/stabilityai/stable-diffusion-3.5-large →

by Stability AI (models); Comfy Org / comfyanonymous (the ComfyUI interface) · released October 2024 (SD 3.5); actively maintained and used through 2026

Fully local, fully free — the open node-graph workflow

Stable Diffusion 3.5 (Stability AI, October 2024, still the current open-weight line in mid-2026) ships in three sizes — Large (8B), Large Turbo (a faster distilled version), and Medium (2.6B, tuned for consumer GPUs) — and runs natively in ComfyUI, the node-based workflow app that is now the de facto standard local interface for it. Instead of one text box, ComfyUI gives you a visual pipeline: chain a base generation into an upscaler, a ControlNet, a LoRA, or a custom model, one node at a time, and save the whole graph to rerun later. Under the Stability AI Community License, both the weights and your generations are free — including commercial use — for any individual, lab, or organization earning under $1M in annual revenue; above that, an Enterprise license is required. For a Heidelberg lab or student project this is the only option here where nothing — no prompt, no image — ever leaves your own machine.

Best for

Open source / self-host

Also useful for

Draft docs

What it does well

Genuinely free, including commercial use, for any student, lab, or startup under the $1M revenue threshold
ComfyUI's node graph gives full pipeline control — chain upscalers, ControlNets, and LoRAs step by step, and save the graph
Runs fully offline/air-gapped — the only tool on this list where no prompt or image ever leaves your machine
The largest open ecosystem of fine-tunes, LoRAs, and shared workflows (Civitai) of any model here

Where it falls short

ComfyUI's node-graph UI is a real learning curve compared with typing a prompt into a text box
Needs a real GPU (8GB+ VRAM for Medium, more for Large) — without a dedicated GPU you're renting cloud compute instead
Out-of-the-box photorealism and prompt-following trail the paid frontier models here — quality gains come from your own fine-tuning effort

Try it in 5 minutes

Generate your first image fully locally

Install ComfyUI, load the default SD 3.5 Medium workflow template, and prompt: "a microscope slide of stained plant tissue cells, macro photography, scientific lighting."

Expected: An image rendered entirely on your own GPU with zero API calls and zero per-image cost — plus a saved node graph you can tweak and rerun.

Pricing

Model weights and local generation are free, including commercial use, for individuals and organizations under $1M annual revenue (Stability AI Community License); above that threshold, a separate Enterprise license is required. ComfyUI itself is free, open-source software — the only cost is your own GPU, or a rented one.

⚠️ Watch out

Free only for organizations under $1M/year revenue (Stability AI Community License), and ComfyUI's node-graph interface has a real learning curve versus a text box.

🔥 Adobe Firefly (Image Model 5) The commercially-safe choice — trained on licensed data, IP-indemnified

🔥

Adobe Firefly (Image Model 5)

free tier firefly.adobe.com →

by Adobe · released 2023 (Firefly); Image Model 5, 2026

The commercially-safe choice — trained on licensed data, IP-indemnified

Adobe Firefly is trained only on Adobe Stock, openly-licensed content, and public-domain material — deliberately, so the output is "commercially safe" with no scraped-web-image rights ambiguity. Its 2026 Image Model 5 adds native 4-megapixel photorealistic output. It integrates directly into Photoshop, Illustrator, and Express (generative fill/expand), which many students and staff already have via a university Creative Cloud license, and every generation carries Content Credentials (C2PA metadata) documenting its provenance. The real differentiator for a German university context: Adobe indemnifies commercial output on its paid plans — i.e. contractually stands behind you if a Firefly image is challenged as infringing — something none of GPT Image 2, Midjourney, or Nano Banana Pro offer at all. The trade-off is creative range and prompt permissiveness: Firefly optimizes for "safe," not "best-looking," and its content filters are noticeably stricter.

Best for

Draft docs

What it does well

Trained only on licensed Stock/public-domain data — the closest thing to a cleared-rights guarantee among these six tools
Deep integration with Photoshop/Illustrator/Express — generative fill and expand inside tools students already use
Automatic Content Credentials (C2PA) attached to every output — useful for academic or press provenance requirements
A genuine free tier exists, unlike Midjourney or Nano Banana Pro

Where it falls short

Base photorealism and creative range trail Midjourney/FLUX.2/Nano Banana Pro — Firefly optimizes for "safe," not "best-looking"
Reported content filtering is stricter than competitors and can block legitimate scientific imagery (anatomy, chemistry)
The exact scope of IP indemnification (which paid tier, what it covers) varies by source — verify Adobe's current terms before depending on it

Try it in 5 minutes

Generate a commercially-safe hero image and check its provenance

In Firefly (web, free plan): "A university lab with students working at a bench, warm natural light, editorial photography style" — then open the Content Credentials panel on the result.

Expected: An image with attached provenance metadata (Content Credentials) — useful when a professor or press office needs to show the image's generation is documented and Adobe-licensed.

Pricing

Free plan (limited monthly credits); Standard $9.99/mo (2,000 credits); Pro $19.99/mo (4,000 credits, access to multiple partner models); Premium $199.99/mo (50,000 credits, studio-scale). Standard-quality image generation is unlimited on every paid plan — credits are only spent on premium features like video.

⚠️ Watch out

IP indemnification is the headline feature, but sources disagree on exactly which paid tier it fully applies to — confirm Adobe's current terms before relying on it for a real commercial deliverable.

🚀 Dokploy Self-hosted Vercel/Heroku alternative — git push, auto-deploy, your own server

🚀

Dokploy

free tier open source dokploy.com →

by Dokploy (open-source project + Dokploy Cloud as the commercial hosted option) · released 2024

Self-hosted Vercel/Heroku alternative — git push, auto-deploy, your own server

Dokploy is an open-source (Apache-2.0) deployment platform you install on your own server with one shell script. It wraps Docker and Traefik behind a clean dashboard: connect a GitHub/GitLab/Gitea/Bitbucket repo, and every push to the configured branch triggers a webhook-driven build and redeploy — no CI pipeline to hand-wire. It natively understands Docker Compose (paste a compose file and it manages the whole multi-container stack, not just one container at a time), issues and renews Let's Encrypt certificates automatically via Traefik once you point a domain at it, and can back up its own database and config to S3. For a lab or small team, it turns a bare Hetzner VPS into something that behaves like Vercel or Heroku — for a fraction of the monthly cost and with your data staying on hardware you control, in the EU if you choose an EU region.

Best for

Automate workflow
Build an app

Also useful for

Write code
Data analytics

What it does well

One-line install (curl script) on any Ubuntu/Debian/Fedora/CentOS box — dashboard live in minutes
Native Docker Compose support — deploy a whole multi-container stack, not one container at a time
Git push → webhook → auto-deploy across GitHub, GitLab, Gitea, Bitbucket, and Docker Hub
Automatic HTTPS via Traefik + Let's Encrypt once a real domain is pointed at it
Low overhead — roughly 350 MB RAM and under 1% CPU idle, so a small VPS goes a long way

Where it falls short

Free tier caps concurrent builds at 2 — a busy team hits this fast
Multi-server support runs on Docker Swarm — workable for simple multi-node setups, less mature than Coolify's multi-provider story
You own the server: OS updates, Dokploy version bumps, and disk space are your responsibility
Backups cover Dokploy's own Postgres DB + config directory, not your deployed apps' data automatically — each app still needs its own backup job

Try it in 5 minutes

Install Dokploy and see the dashboard

curl -sSL https://dokploy.com/install.sh | sh

Expected: A running Dokploy dashboard at http://<server-ip>:3000, ready to create your admin account — proof the control plane is live before you connect any app.

Pricing

Self-hosted core is free forever (Apache-2.0), capped at 2 concurrent builds. An Enterprise license (SSO/SAML, audit logs, custom roles, unlimited builds, whitelabeling) is available self-hosted, price on request. Dokploy Cloud (they run the server for you) starts at $4.50/server/month (Hobby) or $15/month (Startup) — a different product from what this chapter teaches.

Open the full Dokploy course 5 lessons

⚠️ Watch out

You are now the ops team: server patches, Dokploy version upgrades, and backups are your job, not a vendor's. The free tier also caps you at 2 concurrent builds.

💬 Discuss this chapter

Ask, share, or report — over on the Heidelberg AI community forum.

Ask a question Stuck on something? The community can help. Show what you built Share it — others can try it and cheer it on. Report a problem or idea Spotted a bug, or have an improvement?