Langfuse is an LLM engineering platform for the observability and evaluation of AI applications and agents. Observability means recording what a running program does, so that engineers can inspect it later. Langfuse is made by the Langfuse team, which has joined ClickHouse.
An application sends traces to Langfuse through the Python or TypeScript SDK, or through OpenTelemetry, the open standard format for traces. A trace is a hierarchical record of one request: each LLM call, tool call and retrieval step, with its input, output, cost and latency. On top of the traces, Langfuse offers prompt management (prompts stored outside the code, with deployments and rollbacks), evaluation by LLM-as-a-judge (one language model scores the output of another), heuristic functions or human review, experiments on test datasets, and dashboards with alerts for cost, latency and quality.
All product features are open source under the MIT license, and a team can self-host Langfuse with Docker Compose, Kubernetes or Terraform. Langfuse Cloud is the hosted service: a free Hobby plan, then monthly subscriptions that include a number of units and charge for additional units.