Frappe Docs MCP Server
Provides read access to Frappe framework documentation through MCP, enabling search and retrieval of documentation pages via natural language.
README
Frappe Docs MCP Server
A Model Context Protocol server that gives an MCP client read access to the Frappe framework documentation.
It speaks MCP over stdio using the official
@modelcontextprotocol/sdk,
so any stdio-capable client (VS Code, Claude Desktop, Claude Code) can run it.
Tools
search_frappe_docs
Finds documentation pages by keyword. Start here when you do not already know a page's exact path.
| Parameter | Type | Description |
|---|---|---|
query |
string | Keywords, e.g. create doctype |
limit |
number | Optional, 1-50, default 10 |
Every docs page embeds the full navigation tree, so one fetch yields a catalogue of all ~215 pages. Every query token must match somewhere in a result's title or path, so adding a word narrows the results rather than widening them.
1 page matched "create doctype":
- **Create a DocType** — `en/tutorial/create-a-doctype`
get_frappe_doc
Fetches a documentation page and returns its readable content as markdown.
| Parameter | Type | Description |
|---|---|---|
path |
string | Path relative to the docs root, e.g. en/tutorial |
Paths always resolve under https://docs.frappe.io/framework/user. A leading
framework/user is stripped if you include it, .. segments are rejected rather
than resolved, and each segment is percent-encoded, so a path cannot escape that
prefix or smuggle in a query string.
A raw docs page is roughly 960 KB of HTML, most of it navigation. Only the article body is kept and converted to markdown, which brings a typical page down to a few KB.
Configuration
Both settings are optional. A malformed value is reported on stderr and the default is used, so a typo degrades rather than killing the server.
| Variable | Default | Meaning |
|---|---|---|
FRAPPE_DOCS_MAX_BYTES |
100000 |
Ceiling on one response, in UTF-8 bytes. 0 disables truncation. |
FRAPPE_DOCS_CACHE_TTL_MS |
900000 |
How long fetched pages and the catalogue stay fresh (15 minutes). |
Pages and the catalogue are cached in memory, so repeated lookups within a session do not re-fetch. Concurrent requests for the same page share a single fetch.
Requirements
Node.js 20 or newer. There is no node-fetch dependency; the server uses the
global fetch built into Node 20.
Quick start
git clone https://github.com/muthanii/frappe_docs_mcp.git
cd frappe_docs_mcp
npm install
npm run build
npm test
Wiring it into a client
The server reads JSON-RPC from stdin and writes it to stdout, so it must be
launched directly rather than through npm run start (npm prints its own banner
to stdout, which corrupts the protocol stream).
VS Code — .vscode/mcp.json is already set up:
{
"servers": {
"frappe-docs-mcp": {
"type": "stdio",
"command": "node",
"args": ["${workspaceFolder}/dist/server.js"]
}
}
}
Claude Desktop / Claude Code — point at the built entry point:
{
"mcpServers": {
"frappe-docs-mcp": {
"command": "node",
"args": ["/absolute/path/to/frappe_docs_mcp/dist/server.js"]
}
}
}
Run npm run build after any source change so dist/ stays current.
Docker
docker compose run --rm -T frappe-docs-mcp
The image publishes no port. This is a stdio server, not an HTTP service, so it must be run attached with a client owning stdin and stdout.
Development
| File | Role |
|---|---|
src/docs-url.ts |
URL construction and path sanitising |
src/catalog.ts |
Navigation-tree parsing and keyword search |
src/extract.ts |
Article extraction, markdown conversion, truncation |
src/cache.ts |
TTL cache with request coalescing |
src/server.ts |
Tool registration and the stdio transport |
src/*.test.ts |
node:test suites (84 tests) |
Logic is kept out of server.ts so it can be tested without spawning a process.
npm run build uses tsconfig.build.json, which excludes *.test.ts so tests
do not ship in the image. npm test compiles everything, including the tests.
package-lock.json is committed and the Docker build uses npm ci, so image
builds are reproducible.
Diagnostics must go to stderr. Anything written to stdout is interpreted as a protocol frame; CI asserts that the server writes nothing to stdout before a client speaks.
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。