Arize Phoenix
Arize Phoenix is an AI observability platform from Arize for experimenting with, evaluating and troubleshooting AI applications. It traces applications through OpenTelemetry, runs response and retrieval evaluations with models as judges, and keeps versioned datasets, experiments, a playground and prompt management, with an AI engineering agent and an MCP server built in. It is under the Elastic License 2.0, which forbids offering it as a hosted service, so this site labels it source available.
What does Arize Phoenix do?
Phoenix is "Arize's open-source AI observability platform designed for experimentation, evaluation, and troubleshooting", in its README's words, and is "built on top of OpenTelemetry and is vendor, language, and framework agnostic." Tracing uses OpenTelemetry-based instrumentation, evaluations judge responses and retrieval with language models, and versioned datasets, experiments, a playground and prompt management with version control and tagging round out the workflow.
PXI, an AI engineering agent, is built in, and Phoenix exposes a remote MCP server. The README's integration tables list 35 Python, 9 JavaScript, 4 Java and 2 Go integrations and 9 platforms, including OpenAI, Anthropic, Google GenAI, Bedrock, LiteLLM, Groq, Mistral, Ollama, Dify and LangFlow. Code evaluators can run in E2B, Daytona, Vercel Sandbox or Modal. It collects basic web analytics by default, and the README says trace data and evaluation results are never collected.
Key facts
- Licence: the Elastic License 2.0 for the platform. It forbids providing the software to others as a hosted or managed service with substantial access to its features, circumventing licence-key functionality, and removing notices.
- The phoenix-otel and phoenix-client packages are Apache-2.0; phoenix-evals is ELv2. An IP notice says portions are covered by US patents 11,315,043 and 11,615,345.
- Built on OpenTelemetry; vendor, language and framework agnostic.
- Evaluations of responses and retrieval with models as judges; versioned datasets, experiments, a playground and prompt management.
- PXI, an AI engineering agent built in, and a remote MCP server.
- Integrations: 35 Python, 9 JavaScript, 4 Java, 2 Go and 9 platforms, including Ollama, Bedrock, LiteLLM, Dify and LangFlow.
- Install: pip install arize-phoenix then phoenix serve, or uvx arize-phoenix serve, or Docker or Helm. Web analytics are on by default; turn them off with PHOENIX_TELEMETRY_ENABLED=false.
- Arize sells a separate managed product, Arize AX. Latest tagged release when read: arize-phoenix-v20.14.0 on 2026-09-17; 11,523 stars.
What does it replace, and where does it fall short?
Phoenix replaces LangSmith for a team that wants tracing and evaluation built on the OpenTelemetry standard, self-hosted, with integrations for most languages and frameworks and an agent to help analyse traces. It is the pick for OpenTelemetry and language breadth on the open-source alternatives to LangSmith page.
Where it falls short of open source: the Elastic License 2.0 is not OSI-approved and forbids offering Phoenix to others as a hosted or managed service, which is why this site labels it source available, although its README calls it open source. Web analytics are on by default. For an OSI licence, Opik or Langfuse.
How does Arize Phoenix run?
pip install arize-phoenix, then phoenix serve; or run it with no install through uvx arize-phoenix serve. A Docker Hub image and a Helm chart cover containers and Kubernetes. Arize sells a separate managed product, Arize AX; Phoenix itself is self-hosted.
Who is Arize Phoenix for?
A team that instruments with OpenTelemetry and wants evaluation and tracing for its own AI applications, self-hosted. A company that wants to offer observability to its own customers needs an OSI licence such as Opik's.
What limits does the README state?
From the README: the Azure template serves plain HTTP because Azure Container Instances does not terminate TLS; basic web analytics are collected by default and turned off with PHOENIX_TELEMETRY_ENABLED=false.
Questions people ask
Is Arize Phoenix open source?
Not by the OSI definition. The platform is under the Elastic License 2.0, which forbids offering the software as a hosted or managed service and circumventing licence keys, so this site labels it source available, although its README calls it open source. Its OpenTelemetry and client packages are Apache-2.0.
Can I use Phoenix at my company?
Yes. The Elastic License 2.0 allows using and self-hosting it for your own applications. What it forbids is providing Phoenix to third parties as a hosted or managed service with substantial access to its features.
Does Phoenix collect data?
It collects basic web analytics such as page views and interface interactions by default; set PHOENIX_TELEMETRY_ENABLED=false to turn them off. The README says none of your trace data, evaluation results or sensitive information is ever collected.
How does Phoenix compare with LangSmith?
Both trace and evaluate LLM applications. Phoenix builds on OpenTelemetry, lists integrations across Python, JavaScript, Java and Go, and self-hosts under the Elastic License 2.0. The open-source alternatives to LangSmith page compares it with Langfuse and Opik.
How do I start Phoenix?
pip install arize-phoenix and then phoenix serve, or uvx arize-phoenix serve with no install at all, by the README. A Docker Hub image and a Helm chart cover containers and Kubernetes, and the same instrumentation works whether it runs locally or in your cloud.
Sources
- Arize Phoenix README, LICENSE (Elastic License 2.0) and IP_NOTICE: github.com/Arize-ai/phoenix, read 2026-09-17.
Compared with the others
On the open-source alternatives to LangSmith page, Arize Phoenix is the pick for opentelemetry across four languages. Tracing built on OpenTelemetry with evaluations, datasets, experiments and prompt management, plus a built-in agent and an MCP server.
Also on that page: Langfuse for the most widely used, mit core, Opik for one apache-2.0 licence, plus ci tests.
More llm observability and evaluation
An open-core LLM engineering platform, MIT outside its ee folders and part of ClickHouse since January 2026, for tracing, prompt management, evaluations, datasets and a playground, with Python and TypeScript SDKs.
Added September 17, 2026. Every claim above comes from the project's README, LICENSE or model card, read on September 17, 2026, or from the GitHub API on the date shown in the panel. Found an error? Write to reveneau@licheo.com and it is fixed in the next weekly pass. Repository: github.com/Arize-ai/phoenix.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.