MCP server

Host a Model Context Protocol server so coding agents can search and read your docs directly — tools, resources, discovery documents, content-type and facet filters, the Connect to MCP menu, and the server output it needs.

Host a Model Context Protocol server so coding agents (Claude Code, Cursor, VS Code, claude.ai connectors) can search and read your docs directly — no scraping. It’s opt-in:

agents: {
  mcp: {
    enabled: true,
    route: "/mcp", // where the server is mounted
  },
}

If a content or custom page already owns the route, the server isn’t generated: the build warns, and llms.txt and the other discovery documents leave it out.

Add an MCP server to your docs walks through it from turning it on to deploying the server and connecting a client.

Tools and resources

The server exposes read-only tools — search_docs, get_page, list_pages, and get_navigation — and every page as an MCP resource (resources/list enumerates the pages at their served URLs with a text/markdown type, leaving out i18n fallback copies of untranslated pages as list_pages does; resources/read returns the page’s agent Markdown, the same output as get_page), so clients that attach context by URI can browse the docs without calling a tool. It publishes discovery documents at /.well-known/mcp.json and /.well-known/mcp/server-card.json. The server card follows the SEP-2127 Server Card extension schema (reverse-DNS name, remotes transport endpoints), with initialize-shaped compat fields (serverInfo, capabilities, transports) for scanners built against the proposal’s earlier revision. Each page’s Connect to MCP menu offers copy-and-go install for Claude Code, Cursor, VS Code, and Codex (shown once deployment.site is set), and you can choose the clients it lists.

search_docs runs its own full-text index, so it works regardless of your search provider — and even with search: false. The MCP server is a separate feature from on-page search.

Hidden pages stay out of list_pages and resources/list, and out of search_docs unless search.indexing.includeHiddenPages is on, as they stay out of site search and llms.txt. get_page and resources/read still return a hidden page by its route: hidden isn’t private, and the page is still served at its URL and as Markdown.

The same tools are available over plain HTTP as the JSON API, for frameworks that don’t speak MCP.

The endpoint reads a request body only up to 64 KB and answers anything larger with 413. A call to a tool that doesn’t exist, or with arguments that aren’t an object, gets the JSON-RPC Invalid params error (-32602).

Scoping by content type and facets

search_docs and list_pages both accept an optional contentTypes filter, narrowing results to pages of the given frontmatter types — ["rfc"], ["blog", "changelog"] — so an agent working against a site that mixes docs with RFCs, runbooks, or policies can scope retrieval to the kind of page it needs. Every result names its content type, and list_pages output shows the types in use.

Both tools also accept a filters object matching against the facets a site declares per content type (content.types.<type>.facets) — custom frontmatter keys whose values become filterable metadata:

{
  "query": "OpenAPI request schemas",
  "contentTypes": ["rfc"],
  "filters": { "domain": "architecture", "status": "enforced" }
}

Every filters entry must match (results carry their facet values, and list_pages shows each page’s), so a knowledge base can drive progressive-disclosure agent workflows — enumerate the enforced standards, search only within them — without any server of its own.

A value matches exactly, case included, as list_pages shows it: "enforced" doesn’t match a page whose status is Enforced. Facet values are strings, so a number or boolean in filters matches its string form ({"priority": 1} finds priority: 1), and a value of any other shape, such as a list, is ignored rather than filtering.

Connect to MCP menu

The Connect to MCP page action copies the server’s URL, and then lists the clients it installs the server into: commands to copy for Claude Code and Codex, and install links for Cursor and VS Code. clients picks them. true (the default) lists every built-in client, false none, which leaves just Copy server URL in the menu, and a list shows the clients you name, in order, with a rule between the commands and the links.

A client Blume doesn’t ship is an object with a label and the command to copy. {name} in the command becomes the server’s name as an id: name, or the site title, lowercased, with each run of characters other than a–z and 0–9 turned into a hyphen and the ends trimmed (docs if nothing is left). {url} becomes the server’s URL. On a multilingual site, the label can be a map of locale code to text. icon takes a built-in icon name, image path/URL, or inline SVG, and defaults to terminal.

agents: {
  mcp: {
    enabled: true,
    clients: [
      "claude-code",
      "codex",
      {
        label: "Copy Copilot CLI command",
        command: "copilot mcp add --transport http {name} {url}",
      },
      "cursor",
      "vscode",
    ],
  },
}

Server output required

The MCP server is a live endpoint (/mcp), so it can’t run on a static build. Name a host adapter from blume/deploy to switch to server output:

import { node } from "blume/deploy"; // or vercel, netlify, cloudflare

export default defineConfig({
  deployment: node({ site: "https://docs.example.com" }),
});

A static build with agents.mcp.enabled fails fast with a message telling you to set a host adapter. See Deployment for the adapters. Once deployed, connect from Claude Code with:

claude mcp add --transport http my-docs https://docs.example.com/mcp