@particle-academy/docs-mcp
A dev-time MCP server that lets coding agents read documentation directly from installed @particle-academy/\* packages, ensuring version-matched docs without network calls.
README
@particle-academy/docs-mcp
Dev tool — not for production. A local Model Context Protocol server that hands your coding agent the docs shipped inside every installed
@particle-academy/*package. Runs on your machine, talks stdio, exits when your editor closes. Zero runtime dependencies, no network calls, no telemetry.
If you're working on a project that uses react-fancy, fancy-sheets,
fancy-flow, agent-integrations, etc., this lets Claude Code / Cursor
/ Claude Desktop pull from the docs that actually match the versions
you installed — instead of guessing from training data.
What it is (and isn't)
✅ Is: a dev-time MCP server you wire into your editor's MCP config. Spawned as a subprocess when your editor starts, killed when it stops.
❌ Not: a runtime library you ship in your app bundle. Don't import
from it in application code. It's a CLI; the bin is docs-mcp.
❌ Not: a hosted service. Everything runs locally against your own
node_modules and your own packages/ workspace folder. Nothing leaves
your machine.
❌ Not: a docs publisher. It only exposes README.md and docs/**
files that already ship inside @particle-academy/* packages.
❌ Not: a search index — substring grep, no rankings, no embeddings. Good enough for "find me the docs page that mentions X."
Install + configure
You don't install this into your project. Configure your editor to spawn
it on demand via npx:
Claude Code / Cursor (.mcp.json in project root)
{
"mcpServers": {
"particle-docs": {
"command": "npx",
"args": ["-y", "@particle-academy/docs-mcp"]
}
}
}
Claude Desktop (claude_desktop_config.json)
{
"mcpServers": {
"particle-docs": {
"command": "npx",
"args": ["-y", "@particle-academy/docs-mcp"],
"cwd": "/absolute/path/to/your/project"
}
}
}
cwd defaults to the editor's working dir, which is usually correct —
specify explicitly only if the editor launches the server from somewhere
else. The server scans <cwd>/node_modules/@particle-academy/* and (if
present) <cwd>/packages/*.
Restart your editor. The first invocation downloads + caches the package; subsequent launches are instant.
Locally-built / pre-publish
While developing this package itself (or before it's published to npm), point the editor at the local build:
{
"mcpServers": {
"particle-docs": {
"command": "node",
"args": ["/absolute/path/to/packages/docs-mcp/dist/cli.js"]
}
}
}
Run npm run build once in packages/docs-mcp/ before pointing the
editor at it.
Tools exposed to the agent
| Tool | Description |
|---|---|
docs_list_packages |
Every scanned package with name, version, file count, source. |
docs_list |
All doc paths. Optional package filter. |
docs_read |
Read one doc by (package, path). |
docs_search |
Substring search; returns hits with line numbers + section headings. |
docs_refresh |
Re-scan the filesystem (call after npm install of a fancy-* package mid-session). |
All tools return both human-readable text and a structuredContent JSON
payload so agents can either read the formatted lines or destructure rows
programmatically.
Typical agent flow
agent: docs_list_packages
→ 11 packages found
agent: docs_search { query: "controlled component" }
→ 3 hits across react-fancy/docs/Forms.md, fancy-sheets/docs/Spreadsheet.md, ...
agent: docs_read { package: "@particle-academy/react-fancy", path: "docs/Forms.md" }
→ full markdown
CLI flags (for debugging)
docs-mcp [options]
--cwd <dir> Project root to scan from (default: process.cwd())
--scope <name> Restrict to scope(s). Repeatable. Default: @particle-academy
--scope-any Scan every npm scope
--package <name> Scan only these packages. Repeatable.
--include-unscoped Include unscoped packages
--no-workspace Don't scan <cwd>/packages/*/docs even if present
--list Print discovered packages and exit
-h, --help Show this help
Smoke-test the scan without spinning up MCP
npx @particle-academy/docs-mcp --list
Lists every package + doc path the scanner found, then exits. The fastest way to verify it's seeing what you expect.
Smoke-test the MCP loop without an editor
printf '%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
'{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"docs_search","arguments":{"query":"slash command","limit":3}}}' \
| npx @particle-academy/docs-mcp
Three JSON-RPC frames go in, three come back. Useful when debugging an editor config that isn't reaching the server.
How the scan works
For each candidate package, the server includes:
<package>/README.md(if present, exposed at pathREADME.md)- Every
.md/.mdxfile under<package>/docs/**
Paths are stable docs-relative strings (README.md, docs/guides/sheets.md,
docs/api/components.md) — that's what docs_read accepts as the
path argument.
Scan sources, in order:
<cwd>/node_modules/@particle-academy/*/— installed packages.<cwd>/packages/*/— monorepo workspace packages (whenpackages/exists). In-tree packages win over their installed counterparts for the same package name, so docs match the code you're editing.
Subdirectory node_modules are not recursed. Symlinks are followed.
No watchers. If you npm install or pull new docs mid-session, call
docs_refresh (or just restart the editor). Avoids surprising file
events and keeps the process simple.
Privacy
Everything stays on your machine. The process reads markdown files from disk and writes JSON to stdout. It never opens a socket, makes a fetch, or phones home.
Troubleshooting
docs_list_packagesreturns nothing: run with--listfrom the samecwdyour editor uses. If nothing shows up there, the issue is scope/scan — try--scope-anyor--cwd <path>to widen.- Editor says the MCP server crashed: run the CLI directly (
npx @particle-academy/docs-mcp) and paste in a single{"jsonrpc":"2.0",…}line. Errors are written to stderr; the editor usually hides those. - Versions look wrong: in a monorepo, the in-tree
packages/*/package.jsonwins overnode_modules/@particle-academy/*/package.json. That's usually what you want. Pass--no-workspaceto forcenode_modules.
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。