Langfuse

Langfuse is a fuse images tool. Open source LLM observability and evaluation platform that traces every agent step, manages prompts, runs experiments and collects human annotations so teams can improve quality, cost and latency from production data.

Langfuse is an open source platform for tracing, evaluating and improving AI agents and LLM applications. It brings observability, prompt management, datasets, experiments and human annotation into a single workspace, so teams can watch what a model actually does in production rather than guessing from sample outputs. Hierarchical traces capture every LLM call, tool invocation and retrieval step, and each trace can be filtered by user, session, cost, latency or custom metadata. The workflow is built around a loop: launch, observe, improve, repeat. Production traces surface the cases where an agent behaves badly, those cases become dataset items, and experiments compare prompt or model changes against a baseline. Evaluation can run as an LLM judge, as heuristic functions or through human review, and annotation queues route specific traces to reviewers. Prompt management with a playground lets teams edit and version prompts next to the data that shows whether they work. Metrics give a running view of quality, cost and latency so degradation is visible before users report it. The audience is primarily engineering and product teams shipping LLM features: AI engineers instrumenting an agent, platform teams standardising tracing across projects, and quality leads running structured evaluation. OpenTelemetry support and native integrations with frameworks and model providers mean instrumentation usually starts from existing code rather than a rewrite. A free cloud tier and self-managed open source deployment cover both early experiments and stricter data requirements. Against alternatives, Langfuse competes on breadth: many tools cover tracing alone or evaluation alone, while this one connects traces, prompts, datasets, experiments and annotations. That breadth is also the trade off, since a small team may only need a subset. Buyer attention should go to how much instrumentation effort the stack requires and whether the self-managed route matches internal infrastructure capacity.