Tool

Langfuse

Dev AI · LLM observability · Open core

An open-core platform for tracing LLM apps: prompts, generations, scores, and datasets so you can see what the model actually did — self-host the MIT core or use Langfuse Cloud.

What it is

Langfuse is LLM engineering / observability software: SDKs and a UI to capture traces (inputs, outputs, latency, cost signals), manage prompts, and run evals against real traffic or datasets. Teams use it next to gateways like LiteLLM and app platforms like Dify. Docs cover self-hosting and Cloud. Source lives at github.com/langfuse/langfuse.

What it replaces

It replaces “grep logs and hope” / spreadsheet prompt versions when you need structured traces and a shared UI for debugging LLM behavior — closer to product analytics for generations than to a coding agent.

It does not replace the model, the gateway, or a full APM for non-LLM services. It is not a coding agent (Aider, Cline). Hosted Cloud and Enterprise features are product paths on top of the open core.

License

Langfuse is open core. Content outside the enterprise directories is available under the MIT License (“MIT Expat”), verified from the project’s primary LICENSE. Material under ee/, web/src/ee/, and/or worker/src/ee/ is under a separate Langfuse Enterprise License (ee/LICENSE) — commercially licensed periphery; do not assume MIT covers those paths. Langfuse Cloud has its own terms and pricing. Same pattern as other open-core Dev AI tools (e.g. Tabby’s ee/ split).

Where it falls over

  • Self-hosting means Postgres (and related stack) ops, upgrades, retention, and PII — traces often contain user prompts.
  • Enterprise features live under ee/; “we self-host MIT” does not unlock every Cloud checkbox.
  • Instrumentation is only as good as what you send — missing spans and silent SDK misconfig look like “nothing happened.”
  • High-volume apps generate lots of data; storage and sampling policy matter or costs/disk explode.
  • Evals still need human or automated scorers you trust — a dashboard does not invent ground truth.
  • Not a substitute for model choice, prompt skill, or retrieval quality upstream.

Who it's for

Teams shipping LLM features who need shared traces and prompt history — especially beside LiteLLM, Dify, or custom apps — and who will either pay for Cloud or operate the MIT self-host path honestly. Not for “I just want ChatGPT locally” (see the Local LLMs cluster).

Related in this cluster

  • LiteLLM — OpenAI-compatible gateway often sitting in front of providers
  • Dify — LLM app / RAG platform with its own license caveats
  • AnythingLLM — local-first RAG UI (different job: chat product vs observability)
  • Tabby — another open-core Dev AI tool with an ee/ split
  • OpenHands — coding-agent platform; different layer than traces
  • Dev AI hub — agents, RAG, and related tooling

Verdict

Use Langfuse when you need LLM traces and prompt/eval workflow with a clear MIT core you can self-host. Read the ee/ Enterprise License before assuming every feature is free; treat Cloud as a product decision. Pair with a gateway and fix retrieval/prompts upstream — observability shows the mess, it does not invent quality.

← Tools · Dev AI · Home · langfuse.com · GitHub