Skip to content
Arize Phoenix MCP Server

Arize Phoenix MCP Server

by Arize AI
Released
Apache-2.0
Repository Docs npm
Developer Tools

Summary

Official MCP server for Arize Phoenix: read traces and spans, manage prompt versions, and run and inspect datasets and experiments from an AI assistant.

Features

  • List and read traces, spans and span annotations
  • Nine prompt tools including creating and updating prompt versions
  • Manage datasets and add examples from failing spans
  • List and fetch experiments and their results
  • Walk sessions and list annotation configs
  • Works against Phoenix Cloud or a self-hosted deployment

Installation

Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.

Any MCP-compatible agent

Most agents (Claude, Cursor, Windsurf, VS Code, and more) read a standard mcpServers configuration.

  1. Open your agent's MCP configuration file.
  2. Merge the snippet below into it, filling in the environment variables with your own values.
  3. Restart the agent — the "Arize Phoenix MCP Server" tools will be available.
{
  "mcpServers": {
    "arize-phoenix-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@arizeai/phoenix-mcp@latest",
        "--baseUrl",
        "https://app.phoenix.arize.com"
      ],
      "env": {
        "PHOENIX_API_KEY": "<your-phoenix-api-key>"
      }
    }
  }
}

Description

Arize Phoenix is the open-source LLM observability and evaluation platform, and @arizeai/phoenix-mcp is its official Model Context Protocol server. It gives an assistant a single interface to the three things Phoenix stores: the traces your application emitted, the prompts you version, and the datasets and experiments you evaluate against.

The tool surface is unusually broad for an observability server — more than 30 tools across seven areas. Prompts get nine of them, including creating and updating versions, so an agent can iterate on a prompt and commit the result to Phoenix rather than leaving it in a scratch file. Traces and spans can be listed and read, span annotations retrieved, and sessions walked. Datasets support listing, reading, adding examples and enumerating the experiments run against them, and experiments can be listed and fetched by ID.

In practice that makes it a way to close the evaluate-and-improve loop from inside a coding agent: pull the failing spans, look at the prompt version that produced them, add the failures to a dataset as examples, and re-run the experiment — all without leaving the editor.

It runs over npx against either Phoenix Cloud or a self-hosted instance; point --baseUrl at your deployment and supply an API key. Apache 2.0 licensed, developed in the Phoenix monorepo at Arize-ai/phoenix.

Related MCP Servers

MCP: Langfuse

by Langfuse

Langfuse's own MCP server for LLM observability — read traces and observations, manage prompt versions, run datasets and evaluators, and query cost and latency metrics from inside your agent.

Developer Tools

MCP: Weights & Biases MCP Server

by Weights & Biases

Featured

The official W&B MCP server: query experiment runs, Weave LLM traces, artifacts and registries in natural language, and write findings back as a W&B report.

Developer Tools

MCP: Everything

by Model Context Protocol

A reference/test server that exercises every MCP feature — prompts, resources, tools, and sampling — for building and debugging clients.

DevelopersEvaluation
Browse all MCP servers →