Langfuse

An open-source (MIT) LLM engineering platform that integrates tracing, evaluation, prompt management and datasets in one place. Based on OpenTelemetry, it continuously improves AI applications with production data and is self-hostable.

AI Developer Tools#Open Source#Observability#Evaluation

Updated: 2026-10-26

What is Langfuse?

Langfuse is an open-source (MIT) LLM engineering platform, officially positioned as an open platform for "tracking, evaluating and improving AI Agents" (Open Source Agent Evals & Observability). Its approach is to first record real online requests, then use this production data to understand behavior and collaboratively fix issues, thereby delivering higher-quality versions at lower cost and latency. The official summarizes this cycle as "Launch, observe, improve — repeat". The company discloses that 21 Fortune 500 companies use it, processing over 90 billion observation data records monthly and supporting over 100,000 engineers in building on it. It is also described as "Made for developers, loved by agents", with sufficient production validation.

Who is it for?

Engineering teams that deploy LLM applications to production and require an observability and evaluation closed loop; enterprises with high data compliance requirements that prefer self-hosting; and technical teams that need to centrally manage multiple LLM applications and build an organizational-level observability platform.

Key features

End-to-End Tracing

Fully record LLM calls, tool calls and Agent steps, and use Metrics to understand the behavior, latency and cost of each request to quickly locate issues.

Evaluation System

Supports dataset regression, LLM-as-a-judge, and human annotation scoring (Annotations), turning "good or bad" into quantifiable metrics.

Prompt Management

Versioning, comparison and rollback of prompts, decoupled from code; built-in Playground for convenient debugging and experimentation.

Native OpenTelemetry

Collected based on the OTel standard, compatible with 100+ integrations, and the official emphasizes that it works with any stack.

Self-hosted or Cloud

The open-source version can be deployed on your own, keeping your data in your hands; it has async by default, enabling it to support high-throughput production loads.

How to use Langfuse

First, connect the tracking to see the real traffic and failure patterns, then build the evaluation dataset for regression analysis, and only then optimize the prompts – reversing the order will lead to judgments based on noise.
The evaluation cases are recommended to be selected from real failed requests, which are more useful than those created out of thin air.
when collaborating in a team, run a closed loop of annotation, repair and rollback, so that the platform value can truly be realized.
The official also provides an AI assistant (Langfuse Assistant) to assist with integration and troubleshooting.
If you encounter burying point problems, you can take this path first.
even if only a small number of key applications are integrated, you can immediately get a clear visible view of calls and failures.

Things to know

Tracing and evaluation itself has a learning curve; teams need to invest time to get the instrumentation right, otherwise the data will be incomplete.

The cloud version processes call data through third parties; for sensitive scenarios, evaluate the self-hosting solution and data residency clauses.

The quality of evaluation metrics depends on annotation and dataset design; perfunctory evaluations are more likely to mislead decision-making than no evaluations at all.

High-throughput scenarios generate massive observability data, leading to storage costs; pay attention to sampling and retention strategies.

Evaluation and annotation require human input; focus on high-value use cases in the core workflow first, and don’t aim for full coverage right away.

Self-hosted deployment requires maintaining infrastructure; for small team sizes, the free tier of the cloud version may be sufficient.

If the tracing coverage is incomplete, the tracing diagram will miss certain steps. Before troubleshooting issues, verify that all relevant calls have been integrated.

Review evaluation results regularly; indicators that haven’t been updated for a long time will gradually become distorted and fail to reflect the real online performance.