Silica Docs

MCP server

Let external AI agents read your site over the Model Context Protocol, authenticated with API keys.

Besides the built-in AI assistant, Silica can expose your site to external AI agents over the Model Context Protocol (MCP). Claude Desktop, Claude Code, IDE agents, and your own automations can then search your pages and read the original markdown.

The MCP server is off by default and builds on the assistant. A site without assistant.mcp never installs or loads the MCP SDK and has no /api/mcp route.

Enabling the MCP server

  1. Enable the AI assistant first, then install the MCP SDK next to it:
Bash
npm install @modelcontextprotocol/sdk
  1. Turn it on in silica.config.ts:
TypeScript
import { defineConfig } from "@silicajs/core";

export default defineConfig({
  // ...
  assistant: {
    provider: "openai",
    model: "gpt-5-mini",
    mcp: true,
  },
});
  1. Generate one API key per client and put them, comma-separated, in the SILICA_MCP_API_KEYS environment variable:
Bash
npx silica mcp-key
Bash
SILICA_MCP_API_KEYS=slk_...,slk_...

The next silica dev or silica build generates the /api/mcp route. It bypasses the site's sign-in and checks the API key instead, so no browser session is needed. Keys are static: to revoke one, remove it from the variable and redeploy.

Connecting a client

The endpoint speaks stateless Streamable HTTP. Point any MCP client at https://your-site.example/api/mcp with an Authorization: Bearer <key> header.

Claude Code:

Bash
claude mcp add --transport http docs https://your-site.example/api/mcp --header "Authorization: Bearer slk_..."

Claude Desktop and other clients that take a JSON config:

JSON
{
  "mcpServers": {
    "docs": {
      "type": "http",
      "url": "https://your-site.example/api/mcp",
      "headers": { "Authorization": "Bearer slk_..." }
    }
  }
}

What agents get

Every tool is read-only and only sees published pages, exactly like the chat assistant. Drafts and excluded files are invisible (see Drafts and publishing).

ToolWhat it does
search_pagesFull-text search with the same #tag and tag: operators as the site search palette.
read_pageReturns the original markdown of one page by slug or source path.
list_pagesLists every published page with its slug, title, and link.
run_shellRuns read-only shell commands (find, grep, cat, …) in the assistant's content sandbox.

To expose only some of the tools:

TypeScript
export default defineConfig({
  assistant: {
    provider: "openai",
    model: "gpt-5-mini",
    mcp: {
      tools: ["search_pages", "read_page"],
    },
  },
});

Rate limiting

Each API key may make 60 requests per minute by default. Tune it per site, or disable it when another quota guard protects the endpoint:

TypeScript
export default defineConfig({
  assistant: {
    provider: "openai",
    model: "gpt-5-mini",
    mcp: {
      rateLimit: { maxRequests: 120, windowMs: 60_000 },
    },
  },
});