> ## Documentation Index
> Fetch the complete documentation index at: https://arize-ax.mintlify.site/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Arize AX MCP Server

> Give your coding agent read access to Arize AX projects, traces, datasets, experiments, evaluators, monitors, and prompts over the Model Context Protocol.

<Warning>
  The MCP server is in **Alpha**. It is meant for developers exploring new capabilities, and breaking changes are expected. Tool names, arguments, and response shapes can change without a deprecation window. See [API version stages](/docs/ax/rest-reference/overview#api-version-stages) for what each stage means.
</Warning>

The Arize AX MCP server is a hosted [Model Context Protocol](https://modelcontextprotocol.io) server that puts your Arize AX resources inside your agent's context. Connect it once and Claude Code, Cursor, Claude Desktop, Codex, or any other MCP client can look up a project, pull the spans from a failing trace, check what an evaluator scored, or read the current version of a prompt, without you copying IDs between a browser tab and a chat window.

It is a thin layer over the [REST API v2](/docs/ax/rest-reference/overview).

## Connect your client

The server is hosted, so there is nothing to install. You need an endpoint URL and an Arize API key. [Connect your client](/docs/api-clients/mcp/connect) walks through setup for Claude Code, Cursor, Claude Desktop, and Codex.

## What your agent can read

The server exposes 45 read-only tools, plus a `server_info` tool for checking which version you reached.

The [tools catalog](/docs/api-clients/mcp/tool-catalog) has every tool name and its arguments.

## Read-only by design

Nothing the server exposes can change your data. There are no create, update, or delete tools, so an agent that goes off script cannot drop a dataset or overwrite a prompt. Two categories are left out on purpose:

1. Writes and deletes, which need a docstring stating their blast radius before an agent should be trusted to call them.
2. The account administration surface, including API keys, users, roles, and role bindings. Reading those is an information exposure risk even though it changes nothing.

To create or modify resources from an agent, use [Arize Skills](/docs/ax/skills/overview) instead. Skills drive the [Arize AX CLI](/docs/api-clients/cli/overview), which covers the full read and write surface.

## How access works

You authenticate with an Arize API key, passed as a bearer token that gets forwarded to the REST API on every call. See [endpoints](/docs/api-clients/mcp/endpoints) for the regional URLs and [authentication](/docs/api-clients/mcp/authentication) for how to get a key.

## MCP server, skills, or the CLI?

There are three ways to put Arize AX in front of an agent, and each solves a different problem:

* **MCP server** when you want the agent to answer questions about data already in AX. It can debug a trace, check experiment results, or find the project that contains a span. There is no installation step, and the tools remain available across sessions.
* **[Skills](/docs/ax/skills/overview)** when you want the agent to do the work: add tracing to an app, build a dataset, run an experiment, create an evaluator, optimize a prompt. Skills can write.
* **[The Arize AX CLI](/docs/api-clients/cli/overview)** when you are scripting or wiring Arize into CI/CD and no agent is involved.

Using the MCP server and skills together is a reasonable setup. The agent reads through MCP tools and acts through the CLI.

## Next steps

<CardGroup cols={2}>
  <Card title="Connect your client" icon="plug" href="/docs/api-clients/mcp/connect">
    Per-client setup for Claude Code, Cursor, Claude Desktop, and Codex.
  </Card>

  <Card title="Endpoints" icon="globe" href="/docs/api-clients/mcp/endpoints">
    Regional URLs for the MCP server.
  </Card>

  <Card title="Authentication" icon="key" href="/docs/api-clients/mcp/authentication">
    Get an API key and see how bearer token auth works.
  </Card>

  <Card title="Tool catalog" icon="list" href="/docs/api-clients/mcp/tool-catalog">
    Every tool, grouped by resource.
  </Card>

  <Card title="REST API reference" icon="code" href="/docs/ax/rest-reference/overview">
    The API behind the tools, including authorization, errors, and rate limits.
  </Card>

  <Card title="Skills" icon="sparkles" href="/docs/ax/skills/overview">
    Give your agent the write side of Arize AX through the CLI.
  </Card>
</CardGroup>