Langfuse
4.6
Langfuse records what an LLM application actually did — traces, inputs, outputs, cost, latency — and adds evaluation and prompt management on top. It is open source and self-hostable, which puts it in reach of teams whose prompts and outputs cannot be sent to a vendor.
- Commercial use OK
- Free tier
- Enterprise contract
- Pricing
- $29/mo
- Free tier
- Free tier
- Commercial use
- Permitted
- Japanese UI
- Not confirmed
- Enterprise contract
- Available
What it is good for
- Finding out why a production LLM feature behaved differently than in testing
- Tracking cost and latency per feature rather than per account
- Managing prompts outside application code
- Teams that cannot send prompts and outputs to a hosted service
Strengths and trade-offs
Strengths
- Open source and self-hostable
- Traces, evaluation and prompt management in one place
- Cost and latency attribution per trace
- Framework-agnostic SDKs
Trade-offs
- Instrumenting the application is work you have to do
- Self-hosting adds a database and a service to operate
- Evaluation is only as good as the dataset you build
Editors' review
The part of an LLM application you regret not adding until the first incident
LLM features fail differently from ordinary code: no exception, just a worse answer. Without traces you cannot answer why a user got a bad result, which makes improvement guesswork. Langfuse supplies the record.
Cost attribution is the other reason teams install it. 'The OpenAI bill went up' is not actionable; 'this feature's cost per request tripled after the prompt change' is.
Self-hosting is what makes it viable for teams who cannot send prompts and outputs to a vendor — which is a large share of enterprise deployments. Budget for the instrumentation work; nothing here is automatic.
Verdict
Add it before the first production incident rather than after. Self-host if the prompts are sensitive.
Compiled from official documentation and public sources, with our own editorial judgement. We do not publish figures we cannot source.