Local & self-hosted appsPeter Steinberger (@steipete), now maintained by openclaw foundation
openclaw
Decision matrix
openclaw app decision matrix
Overview
OpenClaw is a free, MIT-licensed, self-hosted AI agent that runs on your own machine and talks to you through WhatsApp, Telegram, Discord, Slack, iMessage and dozens of other channels, with persistent memory and full system access (files, shell, web browsing). Pick it when you want a personal agent in your existing chat apps rather than a new dashboard, and you're comfortable with what "full system access" actually means for security.
Other good fits
Self-hosters who want an agent with persistent memory and extensible skills running on their own hardware.
+ Pros
Read pros
OpenClaw is a self-hosted AI agent that runs on your own machine rather than a vendor's cloud, and you talk to it through chat platforms you already use: WhatsApp, Telegram, Discord, Slack, iMessage and 29+ other channels. It has persistent memory across conversations, full system access (browsing the web, managing files, running shell commands), and can be extended with community-built skills and plugins. It's MIT-licensed with no enterprise or paid tier, stewarded by the independent OpenClaw Foundation.
Pick it when…
You want a personal assistant reachable from chat apps you already use daily, not a new interface to check separately.
You want persistent memory and extensibility (skills, plugins) rather than a stateless one-off chat.
You're comfortable self-hosting and paying for your own LLM API usage rather than a bundled subscription.
− Cons
Read cons
Full system access is the point and the risk at the same time: an agent that can browse, manage files and run shell commands on your machine needs the same caution you'd apply to any software with that level of access, more so because it's acting somewhat autonomously based on chat instructions. Community-built skills and plugins extend what it can do, but each one is a new piece of code with its own trust level; review what a plugin actually does before installing it, the same way you would for a browser extension.
Pick something else when…
You want a coding-focused agent instead of a general chat-based assistant.Claude Code or Aider are built specifically for that.
You want the agent to run in an isolated sandbox rather than with full access to your machine.OpenHands or E2B emphasize sandboxed execution.
You'd rather not self-host at all. A hosted assistant removes the machine-management burden, at the cost of not running on your own infrastructure.
Avoid if
You don't have Node.js 24+ or a machine to keep running continuously, plus an LLM API key to pay for.
Workflow
Install OpenClaw on a machine you'll keep running → configure your LLM API key → connect one chat channel to start (don't connect everything at once) → give it a narrow, low-risk first task → review what files it touched or commands it ran → expand to more channels and permissions only once you trust its behaviour on the first one.
Cost
OpenClaw itself is free (MIT license, no enterprise edition). You pay for the machine it runs on continuously, and separately for your chosen LLM provider's API usage (Claude, GPT or another compatible model). There's no bundled subscription; usage cost scales with how much you actually use it, so a chatty daily-driver agent costs more in API calls than an occasional-use one.
Restrictions
OpenClaw needs Node 24 or Node 22.14+, an LLM API key, and a host machine that stays running; there's no managed hosting option since it's self-hosted only. It does offer customizable agent routing and optional sandboxing settings, so full, unrestricted access isn't the only configuration available; review those settings rather than assuming the default. Full system access still means the usual security practices apply beyond that: don't run it as a privileged user beyond what it actually needs, and review third-party skills before installing them.
First project
Documentation-based suggestion — not run here.
Give it one scoped, low-risk task through one channel.
Install OpenClaw (Node 24+ required) and connect it to one chat channel you use regularly.
Configure your LLM API key.
Ask it to do something small and reversible, like summarizing a public webpage or drafting a short reply, not something touching sensitive files.
Review exactly what it did (files touched, commands run) before trusting it with anything more consequential.
Done when you've completed one task and reviewed its actual actions (not just the chat output) against what you expected it to do.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on OpenClaw's homepage, checked 23 September 2026. No installation, channel connection or agent task was performed for this review.
OpenClaw · Sep 23, 2026 — Self-hosted agent concept, chat channels, full system access, persistent memory, skills/plugins, Node.js requirement, MIT license, OpenClaw Foundation.
OpenClaw repository · Sep 23, 2026 — Installation (npm, onboard command), Node 24 or 22.14+ requirement, MIT license, customizable agent routing and sandboxing options.
Run a personal agent that you reach through familiar messaging apps.
Automate email, calendars, web tasks, and shell scripts.
Choose Claude, GPT, or local Ollama models for agent tasks.
Peter Steinberger (@steipete), now maintained by openclaw foundation · free-tier · self-host
personal-agent
messaging-integration
local-llm
persistent-memory
home-automation
email-management
calendar-sync
shell-scripting
automate
talk-think
draft-docs
View sourced imageGateway selection in the official OpenClaw macOS onboarding guide. The source's legacy Clawdbot artwork and blurred nearby-gateway details are retained; this is an application setup screen.
OpenClaw documentation — macOS onboarding
Owner-approved use of public official interface screenshots, as recorded in public-visuals-batch-1.json, 13 September 2026. Copyright retained by OpenClaw authors and contributors; attribution and original redaction retained. Editorial product illustration; no open licence asserted. Converted on 15 October 2026 to a WebP (quality 80) for faster page loads; the original image remains in the media library.
