What it does
Langfuse is an open-source platform that engineering teams use to observe and evaluate LLM applications and agents. Hierarchical traces record each model call, tool invocation and retrieval step, and can be narrowed down by user, session, spend, latency or metadata you attach. Around tracing it adds evaluation (an LLM acting as judge, code-based checks or human reviewers), prompt management with deployments and rollbacks, a playground for testing prompts on real inputs, datasets and experiments, human annotation workflows, and dashboards with alerts for cost, latency and quality. The components work standalone, so tracing can be used without the rest. It is OpenTelemetry native, with Python and TypeScript SDKs and 100+ integrations across agent frameworks, model providers and coding tools. You can use the managed Langfuse Cloud or self-host with Docker Compose, Helm on Kubernetes, or Terraform setups for AWS, GCP and Azure. A CLI, an MCP server, an agent skill and an in-app assistant called Halo are offered for working with Langfuse from the terminal, an IDE or a coding agent.
Capabilities
- Hierarchical tracing of LLM calls, tool use and retrieval steps
- Evaluation with LLM-as-a-judge, custom functions or human review
- Prompt management with versioning, deployments and rollbacks
- Datasets, experiments and a prompt playground
- OpenTelemetry-native, with SDKs and 100+ integrations
- Self-host under the MIT license or use Langfuse Cloud
When to use it
- Debugging and monitoring LLM apps and agents running in production
- Comparing prompt or model changes against datasets before shipping
- Keeping trace and prompt data on your own infrastructure by self-hosting
- Collecting human review of traces to build test datasets
Alternatives to explore
- LangSmith
- Braintrust
- Arize
- Galileo
