# MCP server

Every Readsmith site is an MCP server. An agent that connects gets live tools over the same indexes the site's own search uses: it can search the docs, read any page as Markdown, look up API operations, and flag a page that is wrong. It cannot modify your content or run code, so connecting an agent is safe by construction.

## Connect

The endpoint answers at `/_readsmith/mcp` over the streamable HTTP transport:

```json title="An MCP client entry"
{
  "readsmith-docs": {
    "url": "https://readsmith.dev/docs/_readsmith/mcp"
  }
}
```

That canonical path answers on every host, whether the site is a standalone deploy or a hosted tenant, so it is the reliable form to hand an agent. A standalone deploy also exposes a friendly `/mcp` alias (configurable with `mcp.path`), but it is not guaranteed on a multi-tenant host, so prefer the canonical path when in doubt. The quickest way to get the exact URL for your site is the **Copy MCP URL** action in any page's actions menu, which also offers one-click **Add to Cursor** and **Add to VS Code**.

## The tools

| Tool              | What it does                                                                                                |
| ----------------- | ----------------------------------------------------------------------------------------------------------- |
| `search_docs`     | Hybrid search returning ranked passages with full chunk text, sized for model grounding rather than display |
| `list_docs`       | Every documentation page with its path and description                                                      |
| `get_page`        | One or more pages in full, as Markdown, by path (pass an array to read several at once)                     |
| `list_endpoints`  | The API reference's operations, optionally filtered by tag                                                  |
| `get_endpoint`    | One operation in full: parameters, request and response shapes                                              |
| `submit_feedback` | Report that a page is wrong, outdated, or confusing, so the maintainers can fix it                          |

`search_docs`, `list_docs`, and `get_page` are always available. The endpoint tools appear when the site has an [API reference](/docs/api-reference-guide/setup) configured.

### Flagging a page

`submit_feedback` is the one tool that writes, and it is non-destructive: it records a feedback signal (the same "was this helpful" the reading UI collects), nothing more. The path is validated against the site's real pages, the comment is length-capped, and the unauthenticated endpoint caps how many an agent can send per session. An agent can tell you a page is wrong; it cannot change the page.

## Skill resources

The site's [Agent Skills](/docs/ai/agent-skills) are exposed as MCP resources, so a connected agent discovers and reads them without installing anything; the resource URIs mirror the HTTP paths under `/.well-known/skills/`.

## Limits

MCP requests share the site's rate limiting (60 per minute per IP by default) and the search quality floor: the same retrieval, the same full-chunk grounding.

Tell your users their agents can connect. One MCP entry pointed at your docs beats pasting pages into a context window, and it stays current because it queries the live index.