Langfuse Review: Open-Source Tracing, Evals, and Prompt Management

Who it fits

• LLM apps or agents are in production or close to it. When a run goes wrong, the team needs every model call, tool call, and agent step with its cost and latency.
• Trace data must stay in your own network, or the tracer ships inside a product you sell: the MIT-licensed core self-hosts with all core features and no usage limit, including enterprise SSO. bbv's AI Hub left Arize Phoenix for that license in February 2026, and Langfuse says ClickHouse's purchase brings no planned licensing changes.
• Prompts change often and people outside engineering edit them: labeled prompt versions are fetched and cached at runtime, so a change ships without a deploy.
• The stack mixes frameworks and languages: native integrations cover LangChain, LangGraph, the OpenAI Agents SDK, Pydantic AI, the Vercel AI SDK, and gateways such as LiteLLM, and OpenTelemetry covers the rest.

Where it falls short

• Self-hosting requires ClickHouse, PostgreSQL, Redis, and S3. Docker Compose has no high availability or backups. Practitioners report ClickHouse's memory and CPU use, plus disk growth, as the main burden: one operator measured tuned v4 at 1,848 MiB across four containers and made Phoenix the default on October 13, 2026.
• Cloud bills every trace and observation, plus each score. Deep agent runs and eval scores increase charges. Unrelated OpenTelemetry spans count too.
• Version changes require migrations. v3 moved storage to ClickHouse. On November 16, 2026, Cloud's v4 cutover removes legacy SDKs and APIs, plus trace-level evaluators. Self-hosters pick a backfill that needs about 3x ClickHouse disk headroom, or dual writes until retention rolls over.
• Governance features sit on higher tiers: retention policies need Pro, project RBAC the US$300 Teams add-on, and audit logs Enterprise. Self-hosted, each needs an Enterprise key, now bundled with ClickHouse Cloud, BYOC, or Private and priced on top of that plan.