Langfuse

Langfuse is an open-source LLM observability and evaluation platform that helps engineering teams trace, test and improve AI agents and LLM applications. It covers the loop most teams end up building by hand: capture every step an agent takes, score the results, change a prompt or model, and check that quality improved. Because all product features are MIT licensed and self-hostable, it is a common choice for companies that cannot send production traces to a closed vendor.

What is Langfuse used for?

Langfuse is used across development and production to understand why an LLM app behaves the way it does. Typical uses include:

• Tracing agent runs step by step, including tool calls and retrieval
• Tracking cost and latency per user, session or feature with dashboards and alerts
• Running LLM-as-a-judge or code-based evaluations on production traffic
• Managing prompts outside the codebase with versioned releases and rollbacks
• Building datasets from real traces and running experiments before shipping changes
• Setting up human annotation queues to create golden datasets

Who Langfuse is best for

Langfuse suits AI engineers, ML teams and product engineers shipping LLM features to real users, from startups to large enterprises. Teams with strict data rules value the self-hosted option on Docker, Kubernetes or Terraform. Developers who use several frameworks or call model APIs directly benefit from its framework-agnostic SDKs and OpenTelemetry support. Non-technical teams looking for a no-code agent builder are not the audience.

Langfuse pricing

As of October 2026, Langfuse Cloud has a free Hobby plan with 50,000 units a month, 30 days of data access and 2 users. Core costs $29 per month with 100,000 units, 90 days of data and unlimited users. Pro costs $199 per month with 3 years of data, data retention management, SOC 2 and ISO 27001 reports and higher rate limits. Extra usage is $8 per 100,000 units at first and gets cheaper at volume. A Pro Teams add-on at $300 per month adds enterprise SSO and project-level RBAC. Enterprise starts at $2,499 per month. Self-hosting the open-source version is free.

Langfuse limitations

• Self-hosting at scale means running its database stack yourself
• SSO, SCIM and audit logs sit in higher tiers or add-ons
• It observes and evaluates agents but does not build them, so you still need a framework
• Unit-based billing needs estimating if you log very high trace volumes

Langfuse vs alternatives

LangChain offers its own commercial tracing and evaluation product alongside its framework, which is convenient if your stack is already built on LangChain. LlamaIndex suits teams whose main concern is building retrieval and document pipelines, with observability through integrations. OpenAI Playground is fine for quick prompt testing against OpenAI models but does not trace production agents. Langfuse wins when you want framework-neutral tracing, evals and prompt management under an open-source license you can self-host.

Tips for getting started with Langfuse

Instrument one endpoint first using the SDK or OpenTelemetry and add user and session IDs so traces are easy to filter. Move your most-edited prompts into prompt management before anything else, since rollbacks pay off quickly. Build a small dataset of 20 to 50 real cases and run an experiment whenever you change models. Set cost alerts early.

Langfuse FAQ

Is Langfuse free?

Yes. Langfuse is open source and free to self-host, and Langfuse Cloud has a free Hobby plan with 50,000 units per month. Paid cloud plans start at $29 per month.

Is Langfuse open source?

Yes. Langfuse states that all product features are MIT licensed and can be self-hosted, forked and modified, with deployment options for Docker, Kubernetes and major clouds.

Does Langfuse work with LangChain?

Yes. Langfuse integrates with LangChain as well as Vercel AI SDK, LiteLLM, Pydantic AI, CrewAI and others, and supports OpenTelemetry for other stacks.

What is Langfuse used for?

Langfuse is used to trace LLM applications and agents, monitor cost and latency, run evaluations, manage prompts and compare experiments before release.