Arize Phoenix is Arize's open-source AI observability platform, designed for experimentation, evaluation, and troubleshooting of LLM applications. For users who need to manage production workflows, Arize also offers a hosted platform, Arize AX, and both share the same OpenTelemetry/OpenInference instrumentation.
Key capabilities
- Tracing: Trace the runtime behavior of LLM applications using OpenTelemetry-based instrumentation.
- Evaluation: Use LLMs to benchmark application performance through response evaluation and retrieval evaluation.
- Datasets: Create versioned example datasets for experimentation, evaluation, and fine-tuning.
- Experiments: Track and evaluate changes to prompts, LLMs, and retrieval.
- Playground: Optimize prompts, compare models, adjust parameters, and replay traced LLM calls.
- Prompt Management: Systematically manage and test prompt changes through versioning, labels, and experiments.
- PXI (Phoenix Intelligence): An AI engineering agent built into Phoenix for debugging traces, iterating on prompts, and navigating the product.
- Remote MCP server: Connect MCP clients such as Claude Code and Cursor to a Phoenix instance's /mcp endpoint to query traces, datasets, experiments, and more.
Compatibility Phoenix is vendor-, language-, and framework-agnostic, with out-of-the-box support for major frameworks (OpenAI Agents SDK, Claude Agent SDK, LangGraph, Vercel AI SDK, Mastra, CrewAI, LlamaIndex, DSPy) and LLM providers (OpenAI, Anthropic, Google GenAI, Google ADK, AWS Bedrock, OpenRouter, LiteLLM, etc.). For details on automatic instrumentation, see the OpenInference project.
How to use Phoenix can run on a local machine, in a containerized deployment, or in the cloud. It can be installed via pip or conda, or run without installation using uvx. The fastest way to send traces is to have a coding agent (Claude Code, Codex, Cursor, etc.) instrument the application: run the corresponding command in the project directory, and the setup will detect the framework and LLM provider, install the correct OpenInference instrumentation, and configure trace export. You can also integrate manually in code. The Phoenix container image is available on Docker Hub and can be deployed using Docker or Kubernetes (via Helm chart).
For humans and coding agents Phoenix is designed to be driven by both people and AI coding agents, and provides three interfaces: CLI (npx @arizeai/phoenix-cli), which can retrieve traces, datasets, and experiments and instrument applications; Skills (.agents/skills/), which package workflows to teach agents how to debug, evaluate, and trace with Phoenix; and a remote MCP server, which can connect any MCP client to an instance's /mcp endpoint to query Phoenix directly.
Security and privacy By default, Phoenix collects basic web analytics data (such as page views and UI interactions) to understand usage and improve the product, and does not collect trace data, evaluation results, or any sensitive information. You can opt out of telemetry by setting the environment variable PHOENIX_TELEMETRY_ENABLED=false.
License This software is licensed under the Elastic License 2.0 (ELv2). Portions of the code are protected by one or more U.S. patents.
