Langfuse vs Arize Phoenix: two self-hosted tools for agent traces and evals
Langfuse and Arize Phoenix are both self-hostable tools for tracing and evaluating agents, but Langfuse's core is MIT-licensed open source owned by ClickHouse while Phoenix is source-available under the Elastic License 2.0 and built directly on OpenTelemetry and OpenInference, so licence terms and the hosted upgrade path usually decide between them.
Checked
For a team that wants to run its own tracing and eval stack, Langfuse and Arize Phoenix are the two names that come up first. Both can be self-hosted at no licence cost, both trace model and tool calls, both track cost, and both run evals on datasets and traces.
The licence is the sharpest difference. Langfuse's core is MIT. Phoenix uses the Elastic License 2.0, which is source-available rather than OSI open source. The hosted paths differ too: Langfuse Cloud and ClickHouse commercial plans on one side, Arize AX on the other.
Licence and ownership
Langfuse
Core MIT, with ee/ directories under a commercial licence. Owned by ClickHouse since 16 January 2026, with no planned changes to licensing. 36,761 GitHub stars and v4.72.0 on 29 October 2026.
Arize Phoenix
Elastic License 2.0, which permits self-hosting but bars offering the software to third parties as a hosted or managed service. 12,078 GitHub stars and v21.0.0 on 30 October 2026. OpenInference, its instrumentation layer, is Apache-2.0.
OpenTelemetry
Phoenix describes itself as built on OpenTelemetry and vendor, language and framework agnostic, with auto-instrumentation from the OpenInference project. Langfuse accepts OTLP over HTTP (JSON or protobuf) at /api/public/otel, does not yet accept gRPC, and builds its Python and JS/TS SDKs on the OpenTelemetry client.
Both therefore sit on the same standard. The OpenTelemetry GenAI semantic conventions are still at Development status in their own repository, so attribute names can change between releases.
Evals and cost tracking
Langfuse
LLM-as-a-judge and code evaluators on live traces and datasets, experiments, a GitHub Action for CI regressions, annotation queues and score analytics. Cost per generation from model price definitions or ingested values.
Arize Phoenix
LLM-based response and retrieval evals, versioned datasets, experiments that track prompt, model and retrieval changes, and prompt management. Sandboxed code evaluators arrived in v16.0.0 (May 2026). Cost tracking since v11.0, from token counts and a pricing table, shown per span, trace, session and project.
How to decide
If the legal team requires an OSI-approved licence, or the tool might one day be offered to customers as part of a hosted product, Langfuse's MIT core avoids the Elastic License question. If the team wants instrumentation that is OpenTelemetry and OpenInference from the ground up, and plans to grow into a commercial ML-observability platform, Phoenix fits that path. Either can be trialled on a day of real agent traces before the team commits.
Motiontic's engineers are vendor-neutral on tracing and eval tools. They pick a backend against the client's hosting, data-residency and budget constraints, keep instrumentation portable where the stack allows, build the eval suite from the client's own production failures, and leave the traces, datasets and evals with the client's team.
Langfuse and Arize Phoenix compared for self-hosted agent tracing, as of 1 November 2026
| Langfuse | Arize Phoenix | |
|---|---|---|
| Owner | ClickHouse, since 16 January 2026 | Arize AI |
| Licence | Core MIT (OSI open source); ee/ commercial | Elastic License 2.0 (source-available) |
| GitHub, 30 October 2026 | 36,761 stars; v4.72.0 | 12,078 stars; v21.0.0 |
| Self-hosting | Open-source edition free; Enterprise and Government editions custom | Docker or Kubernetes via Helm |
| Hosted option | Langfuse Cloud, Hobby $0 to Enterprise $2,499 a month | Arize AX, commercial SaaS |
| OpenTelemetry | Accepts OTLP over HTTP; SDKs on the OTel client | Built on OpenTelemetry and OpenInference |
| Evals | LLM-as-a-judge, code evaluators, experiments, CI action, annotation queues | LLM evals, datasets, experiments, sandboxed code evaluators |
| Cost tracking | Per generation, from price definitions or ingested costs | Per span, trace, session and project, since v11.0 |
Common questions
Is Arize Phoenix open source like Langfuse?
Not in the OSI sense. Phoenix is source-available under the Elastic License 2.0, while Langfuse's core is MIT. Both can be self-hosted at no licence cost.
Do Langfuse and Phoenix both support OpenTelemetry?
Yes. Phoenix is built on OpenTelemetry with OpenInference instrumentation, and Langfuse accepts OTLP traces over HTTP and builds its SDKs on the OpenTelemetry client.
What is the hosted version of Phoenix?
Arize AX is Arize's commercial SaaS. Langfuse's hosted version is Langfuse Cloud, with a free Hobby tier and paid plans from $29 a month.
Can Langfuse and Phoenix run on a team's own infrastructure?
Yes. Phoenix publishes container images on Docker Hub and a Helm chart for Kubernetes. Langfuse's open-source edition is free to self-host without limits on core features, with custom-priced Enterprise and Government editions, including air-gapped deployments.
Related
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Langfuse alternatives
Seven tracing and eval platforms compared on licence, hosting, strengths and the teams each one fits.
LangSmith vs Langfuse
Hosted platform from LangChain against the open-source, self-hostable platform now owned by ClickHouse.
OpenTelemetry GenAI semantic conventions
Standard gen_ai.* names for agent and model telemetry, still at Development status.
Trace grading
Scoring a whole recorded agent run, step by step.
Sources
- GitHub: Langfuse licence
- Langfuse: Joining ClickHouse (2026-01-16)
- Langfuse: Cloud pricing
- Langfuse: Self-host pricing
- Langfuse docs: OpenTelemetry
- Langfuse docs: Evaluation overview
- GitHub: Arize-ai/phoenix
- GitHub: Phoenix licence (Elastic License 2.0)
- Phoenix docs: Cost tracking
- OpenTelemetry: GenAI semantic conventions repo
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.