MCP Server

Tiramisu Docs includes a built-in MCP (Model Context Protocol) server that allows AI assistants like Claude, Cursor, and other MCP-compatible tools to browse your documentation programmatically.

The server is hosted as an HTTP JSON-RPC endpoint at /mcp on your deployed docs site. It ships as a SvelteKit route — no CLI, no separate process, no local setup. Deploy your site and the MCP server is live.

Enabling the Server

Set mcp: true in your config to generate a /mcp POST route and show "Connect" buttons in your docs UI:

import { defineConfig } from "@tiramisu-docs/kit";

export default defineConfig({
  mcp: true,
  // ...
});

When enabled, this does two things:

• Generates a /mcp POST route that handles JSON-RPC 2.0 requests from AI tools
• Shows Connect with MCP and Connect to VSCode buttons on your docs pages — the URL is derived automatically from your site's origin

Custom Server URL

If you host your MCP server separately or want the UI buttons to point to a different endpoint, pass a URL string instead:

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

When a string is provided:

• The UI buttons point to the given URL
• No /mcp route is generated — you are responsible for hosting the server at that URL

Connecting AI Tools

Once your site is deployed with mcp: true, AI tools can connect using the hosted URL.

Claude Code

Add to your .mcp.json file in the project root:

{
  "mcpServers": {
    "my-docs": {
      "type": "streamableHttp",
      "url": "https://docs.example.com/mcp"
    }
  }
}

Replace docs.example.com with your actual docs domain.

Cursor

In Cursor's MCP settings, add a new server and paste your docs MCP URL (e.g. https://docs.example.com/mcp). Cursor supports HTTP MCP endpoints natively.

VSCode

Your docs site shows a Connect to VSCode button in the Open dropdown. Clicking it uses the vscode:mcp/install protocol handler to register the server automatically — no manual configuration needed.

Available Tools

The server exposes five tools to AI assistants:

How It Works

The MCP server is a standard SvelteKit POST route at /mcp. It speaks JSON-RPC 2.0 over HTTP — each request is a single POST with a JSON body, and the response is JSON.

Documentation data comes from the virtual:tiramisu-docs module at build time. The server does not scan the filesystem at runtime, which means it works on any deployment target: serverless functions, edge runtimes, static adapters with prerendering, or traditional Node.js servers. Wherever SvelteKit runs, the MCP server runs.