Development & infrastructureLangfuse GmbHOpen source
Langfuse
Decision matrix
Langfuse app decision matrix
Overview
Langfuse records traces of every LLM and tool call in an application, plus prompt versioning, datasets and evaluation scores, so you can inspect and grade behaviour instead of guessing. Pick it once an AI feature is worth debugging systematically. Self-hosting is free and open source; the Hobby cloud tier is also free with 50k units a month, and paid tiers start at $29.
Other good fits
Scoring AI outputs against a defined rubric to catch regressions before they reach users.
+ Pros
Read pros
Langfuse records traces of LLM and non-LLM calls in your application: sessions, users, agent graphs, metrics and alerts, so you can see the exact sequence of calls behind any given output. It also covers prompt management (versioning prompts so you can compare changes), datasets, experiments and evaluation, so you can score outputs against a rubric instead of relying on impressions.
Pick it when…
You want to see the full trace behind a specific AI output, not just the final answer.
You're versioning prompts and want to compare how a change affects behaviour.
You want a defined, repeatable scoring process instead of ad hoc judgement calls on quality.
− Cons
Read cons
Observability helps you inspect what happened; it doesn't prove an answer was correct, safe, or representative of typical behaviour. Scores need a clear rubric behind them, or they just reflect whoever wrote the score's own bias, and a small hand-picked test set won't catch every failure mode. Traces can capture prompts, outputs and identifiers your application handles, so decide your redaction, retention, access-control and residency settings before collecting anything beyond made-up test data.
Pick something else when…
Your application uses one provider and its own telemetry already covers what you need. Provider-native monitoring is one less system to run.
You only need general distributed tracing, not prompt management or evaluation. OpenTelemetry-only tooling is a lighter option.
You need the agent framework itself, not the observability layer.LangChain is what runs the calls that Langfuse would then observe.
Avoid if
Traces will contain sensitive data and you haven't set up redaction or retention limits first.
Workflow
Instrument one application path → send a test request → inspect the resulting trace and its metadata → link that trace to a prompt version or a score → use what you saw to form a specific hypothesis for a change. A human still has to decide whether a trace or a score actually means the application is working well.
Self-hosting is free and open source; run it locally with Docker Compose or deploy to production on Kubernetes. Whichever route you take, this cost is separate from your model-provider bill. Running effort includes deciding what data traces are allowed to contain and setting redaction or retention limits before you send anything sensitive through it.
Restrictions
This review didn't confirm redaction, retention, access-control, residency or self-hosting configuration for any specific project; the documentation describes the capability, not a default that's safe for every dataset. Decide those settings deliberately before sending real (not made-up) data through it.
First project
Documentation-based suggestion — not run here.
Instrument a small FAQ assistant.
Add Langfuse tracing around one prompt in a test application.
Send two fictional questions through it.
Open both traces and read exactly what was sent and returned at each step.
Add a manual score (true/false) for whether each answer actually used the fictional policy text you supplied, rather than inventing something.
Done when you have two traces with a reviewed score each, and you can point to the specific step in the trace that explains why a score was true or false.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Langfuse's documentation (observability, evaluation) and its pricing page, checked 23 September 2026. No SDK, trace, prompt, evaluation, cloud project or self-hosted instance was used for this review.
Trace application runs across models and versions.
Manage prompts and record evaluation scores.
Analyse datasets and results while monitoring model usage.
Langfuse GmbH · free-tier · self-host · cloud
LLM tracing
prompt management
evaluation scores
datasets
cost monitoring
observe
evaluate
automate
View sourced imageExample trace detail from the official Langfuse README. Original interface framing and the publisher's example model output and metrics are retained.
Langfuse — official repository example trace
Owner-approved use of public official interface screenshots, as recorded in public-visuals-batch-1.json, 13 September 2026. Copyright retained by Langfuse and its contributors; attribution 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.