Phoenix
ノートブック内で動作するトレーシングと評価の基盤
概要
OpenTelemetryベースのため、トレースが特定ベンダーに固定されず可搬です。デバッグ中のコードの隣でローカル実行できます。ライセンスはElastic 2.0で、社内利用には問題ありませんが、マネージドサービスとして提供する場合には制約があります。
Phoenixで何ができますか？
- 計装の自動セットアップ —
npx @arizeai/phoenix-cli setup（Phoenix 導入済みならpx setup）がフレームワークと LLM プロバイダを検出し、対応する OpenInference の計装導入からトレース送信の設定までを行います。 - ローカル環境での起動 —
uvx arize-phoenix serveでインストールなしにプラットフォーム全体を起動でき、同じものを Docker Hub のイメージや Helm チャートとしてクラスタへ展開できます。 - 記録した呼び出しの再実行 — Playground では取得済みの LLM 呼び出しをプロンプトやモデル、パラメータを変えて再生でき、その変更はプロンプト管理側でバージョン管理とタグ付けの対象になります。
- 変更を実験として計測 —
arize-phoenix-evalsで応答品質と検索品質を評価し、バージョン付きデータセットに対する実験としてプロンプトや検索の変更を比較できます。 - コーディングエージェントからの照会 — Phoenix 内蔵のリモート MCP サーバーが
/mcpエンドポイントを提供し、Claude Code や Cursor からトレース・データセット・実験を直接読み出せます。
ドキュメント
Arize-ai/phoenix のREADMEより転載（Elastic-2.0）。 原文を読む ↗
Phoenix is an open-source AI observability platform designed for experimentation, evaluation, and troubleshooting. It provides:
- Tracing - Trace your LLM application’s runtime using OpenTelemetry-based instrumentation.
- Evaluation - Leverage LLMs to benchmark your application’s performance using response and retrieval evals.
- Datasets - Create versioned datasets of examples 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- Manage and test prompt changes systematically using version control, tagging, and experimentation.
- PXI (Phoenix Intelligence) - An AI engineering agent built into Phoenix for debugging traces, iterating on prompts, and navigating the product.
- Remote MCP Server - Connect Claude Code, Cursor, and other MCP clients directly to your Phoenix instance’s
/mcpendpoint to query traces, datasets, experiments, and more.
Phoenix is vendor and language agnostic with out-of-the-box support for popular 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, and more). For details on auto-instrumentation, check out the OpenInference project.
Phoenix runs practically anywhere, including your local machine, a containerized deployment, or in the cloud. See Environments for a walkthrough of each option, or jump straight into the Tracing Quickstart.
Table of Contents
- Run Locally
- Trace Your Application
- Deploy
- Packages
- Tracing Integrations
- Sandboxes
- For Humans and Coding Agents
- Security & Privacy
- Community
Run Locally
Install Phoenix via pip or conda and have a fully functional Phoenix. For all installation and hosting options, see the install guide.
pip install arize-phoenix
phoenix serveOr run it with no install using uvx:
uvx arize-phoenix serveTrace Your Application
The fastest way to send traces is to let your coding agent (Claude Code, Codex, Cursor, and others) instrument your app. From your project directory, run:
npx @arizeai/phoenix-cli setup
# or, with Phoenix installed: px setupSetup detects your framework and LLM provider, installs the right OpenInference instrumentation, and wires up trace export. Prefer to wire it up in code? See the tracing documentation.
Deploy
Phoenix container images are available via Docker Hub and can be deployed using Docker or Kubernetes via the Helm chart.
For Docker Compose, Kubernetes/Helm, and other deployment options, see the self-hosting documentation.
[!NOTE] The Google Cloud button builds Phoenix from source in Cloud Shell rather than deploying the prebuilt Docker Hub image. The Azure template serves plain HTTP (Azure Container Instances does not terminate TLS) — front it with a TLS proxy such as an Application Gateway before production use.
Packages
The arize-phoenix package includes the entire Phoenix platform. However, if you have deployed the Phoenix platform, there are lightweight Python sub-packages and TypeScript packages that can be used in conjunction with the platform.
Python Subpackages
| Package | Version & Docs | Description |
|---|---|---|
| arize-phoenix-otel | Provides a lightweight wrapper around OpenTelemetry primitives with Phoenix-aware defaults | |
| arize-phoenix-client | Lightweight client for interacting with the Phoenix server via its OpenAPI REST interface | |
| arize-phoenix-evals | Tooling to evaluate LLM applications including RAG relevance, answer relevance, and more |
TypeScript Subpackages
| Package | Version & Docs | Description |
|---|---|---|
| @arizeai/phoenix-otel | Provides a lightweight wrapper around OpenTelemetry primitives with Phoenix-aware defaults | |
| @arizeai/phoenix-client | Client for the Arize Phoenix API | |
| @arizeai/phoenix-evals | TypeScript evaluation library for LLM applications (alpha release) | |
| @arizeai/phoenix-mcp | Standalone stdio MCP server for older Phoenix versions (maintenance mode — superseded by the remote MCP server built into Phoenix) | |
| @arizeai/phoenix-cli | CLI for fetching traces, datasets, and experiments for use with Claude Code, Cursor, and other coding agents |