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

# MCP server

> Drive Speedtrain from an AI agent — what the MCP server can do, and the boundary it deliberately does not cross.

Speedtrain exposes a **Model Context Protocol** server, so an AI agent can author task configurations, run test sessions, and read results without a person clicking through the UI.

To set it up, click **Use Speedtrain in Claude** in the bottom-right corner of any Speedtrain page, or see [Connect to the MCP server](/speedtrain/mcp-server/connect).

<Info>
  The MCP server is for **building and inspecting**. It can create a configuration, run a preview, and read every score and piece of feedback — but it cannot accept content or ship it. Those decisions stay with people, in the Speedtrain UI.
</Info>

## What it is good at

<Columns cols={2}>
  <Card title="Authoring configurations" icon="pen-to-square">
    Create and push versions from a document, using the platform's own authoring guide as reference.
  </Card>

  <Card title="Inspecting results" icon="magnifying-glass">
    Query review items, read per-rule scores and feedback, and find which rule is doing the rejecting.
  </Card>

  <Card title="Running test sessions" icon="flask">
    Preview a draft version over a selection of documents before anyone publishes it.
  </Card>

  <Card title="Managing destinations" icon="paper-plane">
    List, create, and set the status of destinations.
  </Card>
</Columns>

## What it deliberately cannot do

<Warning>
  * **Accept or decline content.** Every human decision stays in the UI.
  * **Configure tool capabilities.** Those need platform-side credentials and are set up in the UI. A configuration's existing capability entries are carried through untouched when an agent edits the document.
  * **Write documents.** Index documents are read-only over MCP.
</Warning>

That boundary is the point. An agent can do the laborious part — drafting rules, reading hundreds of rejections, spotting which check is noisy — while the decisions that carry consequences remain with a person.

## Getting oriented

Two things are worth knowing before an agent calls anything:

<Steps>
  <Step title="Nothing is inferred from your credential">
    Every call takes a tenant and an environment. The first two calls in any session get them.
  </Step>

  <Step title="The authoring guide is a tool, not a document">
    The server serves its own complete configuration reference and a fill-in template. An agent should read them before authoring or before interpreting classifier feedback, rather than guessing at the document shape.
  </Step>
</Steps>

<Columns cols={2}>
  <Card title="Connect to the MCP server" icon="plug" href="/speedtrain/mcp-server/connect">
    Add the server to Claude, Claude Code, or a script.
  </Card>

  <Card title="Management MCP tools" icon="wrench" href="/speedtrain/mcp-server/tools">
    Every tool the server exposes, and what it takes.
  </Card>
</Columns>