Aller au contenu principal

Connect a reader MCP for visitors

Pro

The MCP (Model Context Protocol) server turns your docs into a live API for AI assistants. A developer using Claude Code or Cursor can connect your docs and the AI can search, fetch, and list pages on demand — far more useful than relying on whatever the AI has scraped.

What gets exposed

Three tools:

  • search_docs(query) — keyword search across every published page. Returns page titles, slugs, and a snippet around the match.
  • get_page(slug) — full markdown of a page by slug.
  • list_pages(category?) — every page on the site, optionally filtered by category.

That's it. Three tools cover 99% of how an AI uses docs in practice.

What it looks like to a visitor

A developer adds your MCP endpoint to their Claude Code config:

{
  "mcpServers": {
    "yourcompany-docs": {
      "url": "https://docs.yourcompany.com/api/mcp"
    }
  }
}

From then on, every Claude Code conversation has live access to your docs. The user types: "set up a webhook for new orders" — Claude searches your MCP, pulls the relevant page, and answers grounded in your real documentation.

Enabling it

Project settings → MCP → toggle on. The endpoint goes live with your next publish. Pro plan only.

The settings page shows your endpoint URL, copy-paste-ready config snippets for popular AI clients, and analytics on which queries hit your MCP server (so you know which docs your AI-using customers reach for).

Who actually uses it

Today, mostly developers using Claude Code, Cursor, Windsurf, and a few AI agent frameworks. Adoption is small but it's the highest-signal users you'll have — they're solving real problems with your product, in their editor, in real time.

This is also where AI agent companies (Vercel AI, LangChain, etc.) tend to look first when building integrations. Having an MCP server means partnerships are one HTTP request away, not weeks of "can we agree on an API spec".

What you can't do (yet)

  • No write access from the reader MCP. It's read-only. (For write access, see the author MCP, which is for editing your own docs from an AI agent.)
  • No subscription/pay-per-use gating from the MCP. It mirrors public access — if your site is publicly readable, the MCP is too. Password-protected sites expose a password-protected MCP.
  • No webhooks on doc updates. When pages change, the MCP serves the new content immediately, but you can't push an update to subscribed clients. (Yet.)

The metric to watch

Project settings → Analytics → "MCP queries" tab. Top queries that hit your MCP, click-through rate to your live docs, and which assistants are calling in (Cursor, Claude Code, raw Claude API, etc.). After a few weeks of having it on, you'll see which parts of your docs are AI-discovery hotspots — and what to write more of.