Skip to content
HUIAHUB

Evaluation & observability

Langfuse

Open-source tracing, evaluation and prompt management for LLM apps and agents, as a cloud service or self-hosted.

Verified Open source
Visit official site (opens in a new tab)

What Langfuse does

Langfuse records what LLM applications and agents do. Each trace captures model calls, tool calls and retrieval steps in a hierarchy, and you can filter traces by user, session, cost, latency or custom metadata. Dashboards and alerts track cost, latency and quality over time.

On top of tracing it adds evaluation and prompt work. Evaluators can be LLM-as-a-judge, heuristic functions or human reviewers in annotation queues, run on production data or in experiments over datasets. Prompt management keeps prompts out of the code, with versions, deployment labels and rollbacks, and a playground tests prompts against real inputs and compares models.

Instrumentation uses native Python and TypeScript SDKs or OpenTelemetry for other languages, and the homepage lists more than 100 integrations, including LangChain, Vercel AI SDK, LiteLLM and the main model providers. A CLI, an MCP server and a ready-made agent skill let coding agents query traces and manage prompts.

The repository is MIT licensed except for its ee folders. Self-hosting the open-source version is free with unlimited usage and runs on Docker Compose, Kubernetes with Helm or Terraform on AWS, GCP and Azure; it needs PostgreSQL, ClickHouse, Redis or Valkey, and S3-compatible storage. Langfuse Cloud is the hosted option. It suits engineering teams running LLM features in production who want to debug, evaluate and version prompts in one tool.

  • Tracing agent runs to find failed or slow steps.
  • Tracking token spend, cost and latency per user or feature.
  • Running LLM-as-a-judge or human evaluations on production traces.
  • Versioning prompts and rolling back a bad deployment.
  • Building regression datasets and comparing experiments.

Who should use it

  • Engineering teams running LLM features or agents in production.
  • Teams that want to self-host their LLM observability data.

Who should look elsewhere

  • Teams without the capacity to run PostgreSQL, ClickHouse, Redis and object storage, unless they use the cloud.
  • Traditional application monitoring that has no LLM component.

Strengths

  • MIT-licensed core that can be self-hosted free with unlimited usage.
  • Tracing, evaluation, prompt management and experiments in one product.
  • Works with any OpenTelemetry-instrumented stack plus native Python and TypeScript SDKs.
  • Cloud runs in separate US, EU and Japan regions.

Limitations

  • The free Hobby cloud plan keeps 30 days of data, allows 2 users and caps ingestion at 1,000 requests a minute.
  • Self-hosting requires four backing services: PostgreSQL, ClickHouse, Redis or Valkey, and S3-compatible storage.
  • Audit logs, SCIM, data retention policies and project-level RBAC need a paid Enterprise license when self-hosted.
  • On Langfuse Cloud, enterprise SSO enforcement and fine-grained RBAC need the $300 per month Teams add-on.

Privacy notes

Langfuse's security page says the cloud service is covered by SOC 2 Type II and ISO 27001 audits and is GDPR compliant. It runs in separate US, EU and Japan regions, and the US region is HIPAA-ready.

Sources