Langfuse

Langfuse is an open-source platform for building, testing, and monitoring LLM applications and AI agents. It grew out of a problem its founders kept hitting while building agents themselves during Y Combinator in 2023: LLM apps are easy to demo and hard to run in production, because the output is non-deterministic, debugging works differently than with normal software, and the iteration loop is messy. Langfuse set out to give engineering teams the tracing and evaluation primitives they needed to see what their AI was actually doing. In early 2026 the company was acquired by ClickHouse, the analytics database it was already built on, though it remains open source and self-hostable.

The platform's foundation is observability. Langfuse traces every step of an LLM or agent workflow, capturing prompts, responses, token usage, cost, and latency, and nesting those steps so a multi-step chain or agent run reads as a structured tree rather than a flat log. Teams use this to debug why an agent took a wrong turn, find slow or expensive calls, and understand behavior across sessions.

On top of tracing, Langfuse adds evaluation and prompt management. Teams can score outputs using language-model judges, code-based checks, or human annotation, build datasets from real production traffic, run experiments to compare changes, and track quality metrics over time to catch regressions. Prompt management lets teams version and deploy prompts separately from application code, so a wording change does not require a redeploy.

A defining trait is its open-source, developer-first posture. Langfuse is MIT-licensed and can be self-hosted at production scale, which appeals to teams with or open-source requirements, and it is framework-agnostic with broad SDK and OpenTelemetry support, so traces can flow to other tools rather than locking teams in. Because it is built on a high-performance analytical database, it is designed to handle the very large volumes of trace data that production and agentic systems generate. In practice, Langfuse serves as the feedback loop that turns opaque model behavior into queryable, improvable data.