mcp-starlight
Enables MCP clients to search and retrieve documentation from an Astro Starlight docs site by reading a pre-built catalog. Supports locale and version filtering, as well as structured data collections.
README
@stellayazilim/mcp-starlight
Give any Astro Starlight docs site an MCP server.
No hosting. The integration writes a catalog into your build; the server runs
on the reader's machine via npx and fetches it. A static host only ever has to
answer GET for a JSON file — which is the one thing it does. GitHub Pages,
S3 and plain nginx all work unchanged.
Locale- and version-aware, and extensible with structured data of your own.
Install
npm install @stellayazilim/mcp-starlight
// astro.config.mjs
import { defineConfig } from 'astro/config';
import starlight from '@astrojs/starlight';
import starlightMcp from '@stellayazilim/mcp-starlight';
export default defineConfig({
site: 'https://example.com',
integrations: [
starlight({ title: 'My Docs' }),
starlightMcp({ siteLabel: 'My Product' }),
],
});
Build the site. You now have dist/mcp-catalog.json, published with everything
else.
Use
Point any MCP client at the published catalog — or at the site root, and the filename is appended for you:
{
"mcpServers": {
"my-docs": {
"command": "npx",
"args": ["-y", "@stellayazilim/mcp-starlight", "https://example.com"]
}
}
}
Claude Code:
claude mcp add my-docs -- npx -y @stellayazilim/mcp-starlight https://example.com
While writing docs, point it at a local build instead:
npx @stellayazilim/mcp-starlight ./dist/mcp-catalog.json
Tools
| Tool | Answers |
|---|---|
search_docs |
"Which page covers X?" — ranked matches with routes and descriptions |
get_doc |
"Show me that page" — full Markdown |
list_docs |
"What is in these docs?" — every page with title and description |
search_reference |
Exact names and signatures from your structured data — only registered when you provide some |
search_docs and list_docs take locale and version when your site has
them, so an answer can be scoped to one language or one documentation line.
Options
starlightMcp({
siteLabel: 'My Product', // shown in tool descriptions
versions: ['preview', 'v1'], // route segments that denote a docs line
exclude: ['api'], // content directories to leave out
filename: 'mcp-catalog.json', // output name inside the build
collections: { … }, // structured data — object or function
})
versions turns a path segment into a facet. With versions: ['preview'],
/preview/guides/x is tagged preview and /guides/x is the default line, so a
reader can ask about one without the other bleeding in.
exclude matches path prefixes after the locale and version segments are
stripped, so exclude: ['api'] covers api/, tr/api/, preview/api/ and
tr/preview/api/ in one entry instead of four.
locales need no configuration — they are read from the Astro i18n config
that Starlight generates.
Structured collections
Prose search cannot answer "what is the exact signature of this type". If your build already produces that information — a generated API surface, a schema, a changelog — hand it over and it becomes searchable alongside the pages:
starlightMcp({
collections: () => ({
api: {
label: '.NET API reference',
entries: types.map((t) => ({
id: t.uid,
title: t.name,
signature: t.declaration,
summary: t.summary,
route: t.docPath,
})),
},
}),
})
Entries need id and title; signature, summary, detail and route are
used when present. The integration knows nothing about where they came from — it
publishes them, and search_reference appears.
Pass a function when the data is produced by another build step, so it is read at build time rather than when the config is loaded.
Why not the Pagefind index?
Starlight already builds one, and reusing it is a fair instinct. It is the wrong shape: Pagefind indexes rendered pages for browser-side search and ships a WASM runtime that returns HTML excerpts. This catalog is Markdown with frontmatter intact, plus whatever structured data you added — which is what a model can actually use, and what makes locale and version filtering possible.
Why a process at all?
MCP is JSON-RPC. A client POSTs a body and expects an answer computed from it; a static host returns the same bytes for every request and rejects POST outright. So something must run — but it does not have to be hosted. Here it runs where the reader is:
your static host the reader's machine
──────────────── ────────────────────
mcp-catalog.json ──GET──► npx …mcp-starlight ──stdio──► Claude, Cursor, …
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。