Langfuse

Key features

• OpenTelemetry-native LLM tracing OpenTelemetry-native LLM tracing: capture hierarchical traces of model calls, tool use, and retrieval steps from any OTel-compatible framework, not only the first-party SDKs.[1][2]

• Session, user, cost, and latency tracking Session, user, cost, and latency tracking: follow multi-turn sessions and per-user behavior with filters and dashboards that expose spend and slow steps.[1][2]

• Three evaluator types Three evaluator types: score outputs with LLM-as-a-judge checks, deterministic code rules, or human annotation queues, all tracked against the same traces.[1][2]

• Versioned datasets and experiments Versioned datasets and experiments: freeze evaluation sets, then run repeatable side-by-side comparisons of prompts and models before rollout.[1][2]

• Prompt management and playground Prompt management and playground: version prompts centrally and test changes in a playground linked back to the traces they produced.[1][2]

• 100+ integrations 100+ integrations: connect frameworks such as LangChain and the Vercel AI SDK plus major model providers through documented integrations.[1][2]

• Public API with Python and TypeScript SDKs Public API with Python and TypeScript SDKs: read and write traces, scores, and datasets programmatically through a documented REST API with an OpenAPI spec.[1][2]

• Cloud or self-hosted deployment Cloud or self-hosted deployment: start on managed Langfuse Cloud or run the same codebase on operator infrastructure with Docker Compose, Kubernetes Helm charts, or AWS, Azure, GCP, Render, or Railway guides.[1][2]

Who will love it

• Teams debugging LLM apps in production: hierarchical traces with session, cost, and latency context isolate failing tool and retrieval steps fast.[1][2]

• Teams building eval loops: datasets, three evaluator types, and experiment comparison turn one-off prompt tweaks into repeatable benchmarks.[1][2]

• Organizations needing deployment choice: managed Cloud tiers with US, EU, or JP data hosting on paid plans, or self-hosting on operator infrastructure, cover opposite compliance postures.[1][2]

Who should avoid it

• Non-technical teams wanting turnkey AI: instrumentation, API keys, and deployment or plan setup assume engineers who own code and infrastructure.[1][2]

• Teams needing fully open-source licensing: the open core is MIT, but enterprise-only folders sit under a separate commercial license, so license-sensitive adopters must check the boundary.[1][2]

• Budget-capped teams ignoring usage metering: Cloud plans add $8 per 100k usage units in overage, and self-hosting shifts cost to infrastructure and operations instead of removing it.[1][2]

Alternatives and compatibility