Summary
Open-source LLM observability — trace, evaluate, and improve AI applications with production data.
Description
Langfuse is an open-source platform for observing and improving LLM applications. When an agent gives a bad answer, Langfuse shows you exactly which prompt, retrieval, and tool call produced it.
What it provides
- Tracing. Nested traces of every model call, retrieval step, and tool invocation, with latency and token cost attributed per step. Framework-agnostic, with SDKs for Python and JavaScript and integrations for the common agent libraries and OpenTelemetry.
- Prompt management. Version prompts, deploy them without redeploying your application, and compare performance between versions.
- Evaluation. Model-based scoring, human annotation queues, and custom evaluators, run against datasets built from real production traces.
- Datasets. Promote interesting production traces into regression test sets — the loop that actually improves quality over time.
- Cost and usage analytics broken down by user, feature, and model.
Deployment
Self-host the whole thing under an open licence, or use the managed cloud. Self-hosting is the reason many teams choose it: traces contain user data, and keeping them in your own infrastructure removes a compliance conversation.
Free tier on cloud, usage-based paid plans, and enterprise self-hosted licensing.
Reviews
Similar App Suggestions
Firecrawl
Turn any website into clean markdown or structured JSON — crawling, JavaScript rendering, and extraction in one API.
Hebbia
AI built for the rigour of finance — run structured analysis across thousands of documents with auditable answers.
Apify
A marketplace of thousands of ready-made scrapers and automation tools, plus the cloud to run your own.
Tavily
A search API purpose-built for RAG and agents — returns synthesised, cited content instead of a list of links.
Exa
A search engine built for AI — embeddings-based retrieval that finds pages by meaning, not keyword overlap.