Langfuse alternatives
What Langfuse is built around, where that leaves voice agents, and which tools fit the jobs it was not shaped for.
Langfuse is an open-source platform for tracing and evaluating AI applications. It is good at what it is built around, and what it is built around is a trace.
That shapes everything else. If your application is a chain of model calls, a trace is the right unit. If it is a phone call, it is not.
Two things to know first
ClickHouse acquired Langfuse in January 2026. Langfuse says it stays open source and self-hostable with no planned licensing changes, and the repository's copyright line now names ClickHouse.
The licence is MIT with a carve-out. The repository excludes its enterprise folder, and the self-hosting page lists features kept for the enterprise tier, including role-based access control, data masking and audit logs.
Where audio sits
Langfuse supports audio as an attachment on a trace, and ships integrations for LiveKit and ElevenLabs. There is no audio judge, no acoustic metric and no telephony.
For an agent whose failures are interruptions, dead air and people hanging up, an attachment on a trace is storage rather than measurement.
What to look at instead, by job
| Built around | Audio | Licence | |
|---|---|---|---|
| Langfuse | A trace | Attachment on a trace | MIT, enterprise folder excluded |
| Coval | A simulated run | Audio-native, places real calls | Not open source |
| Confident AI | A test case | Dedicated voice evaluation API | DeepEval library is Apache 2.0 |
| Evidova | A production conversation | Read from the platform's own audio | Not open source |
- Rehearsing a voice agent before release. Coval simulates calls over real telephony and has a judge that listens to the recording rather than the transcript.
- Writing evaluations as code, in a test suite. DeepEval is Apache 2.0 and runs like Pytest, with a voice evaluation API and named audio metrics.
- Scoring production conversations continuously. That is what we do, and the next section is us arguing our own case, so read it that way.
Where we fit, and where we do not
Evidova scores conversations your agents already had, on the stack you already run, and publishes how far our scores agree with your own reviewers.
| Langfuse | Evidova | |
|---|---|---|
| Before release | Datasets and experiments | Nothing. We only read production. |
| In production | Traces, with online evaluation | Every conversation, scored |
| Self-hostable | Yes | No |
| Open source | Yes, with a carve-out | No |
| Judge agreement | Not published for its own judges | Measured per workspace and published |
If you want an open-source platform you can run yourself, we are the wrong answer and Langfuse is a good one. If you want somebody outside your stack to grade it and show their working, that is the case we make.
Everything above was read on each company's own site and documentation on 5 October 2026. Pricing and features move. Check the source before you make a decision on it, and tell us if we have something wrong.
This page as markdown