codewiki-mcp
MCP server for codewiki.google — AI-powered wiki documentation for open-source repositories. Search repos, fetch wiki content, and ask questions about any repo.
README
<div align="center">
📚 codewiki-mcp
MCP server for codewiki.google — search, fetch docs, and ask questions about any open-source repo
🇷🇺 Русский | 🇬🇧 English
<br />
MCP server that connects any AI assistant to codewiki.google — AI-generated wiki documentation for open-source repositories.
</div>
📖 Overview
codewiki-mcp is a Model Context Protocol server that gives AI assistants access to codewiki.google — a service that generates comprehensive wiki documentation for any GitHub repository. Search repos, fetch full docs, or ask natural-language questions — all through MCP.
✨ Features
| Feature | Description |
|---|---|
| 🔍 Search Repos | Find repositories indexed by codewiki.google |
| 📄 Fetch Wiki Docs | Get full markdown or structured pages for any repo |
| 💬 Ask Questions | Natural-language Q&A with conversation history |
| 🧠 NLP Repo Resolution | Type naturally — wink-nlp extracts keywords and resolves to owner/repo |
| 📡 Multiple Transports | stdio (default), Streamable HTTP, SSE |
| 🔄 Retry with Backoff | Automatic retries with exponential backoff on 5xx errors |
| 🐳 Docker Support | Multi-stage Alpine build |
| 📊 Response Metadata | Byte count and elapsed time on every response |
🚀 Quick Start
Using npx (no install)
npx -y codewiki-mcp@latest
From source
git clone https://github.com/izzzzzi/codewiki-mcp.git
cd codewiki-mcp
npm install
npm run build
Transports
# stdio (default)
node dist/cli.js
# Streamable HTTP
node dist/cli.js --http --port 3000
# SSE
node dist/cli.js --sse --port 3001
🐳 Docker
docker build -t codewiki-mcp .
# stdio
docker run -it --rm codewiki-mcp
# HTTP
docker run -p 3000:3000 codewiki-mcp --http
# with environment variables
docker run -p 3000:3000 \
-e CODEWIKI_REQUEST_TIMEOUT=60000 \
-e CODEWIKI_MAX_RETRIES=5 \
-e GITHUB_TOKEN=ghp_your_token \
codewiki-mcp --http
🔧 MCP Client Configuration
<details> <summary><b>Cursor</b></summary>
Add to .cursor/mcp.json:
{
"mcpServers": {
"codewiki-mcp": {
"command": "npx",
"args": ["-y", "codewiki-mcp@latest"]
}
}
}
</details>
<details> <summary><b>Claude Desktop</b></summary>
Add to claude_desktop_config.json:
{
"mcpServers": {
"codewiki-mcp": {
"command": "npx",
"args": ["-y", "codewiki-mcp@latest"]
}
}
}
</details>
<details> <summary><b>Claude Code</b></summary>
claude mcp add codewiki-mcp -- npx -y codewiki-mcp@latest
</details>
<details> <summary><b>Windsurf</b></summary>
Add to your Windsurf MCP config:
{
"mcpServers": {
"codewiki-mcp": {
"command": "npx",
"args": ["-y", "codewiki-mcp@latest"]
}
}
}
</details>
<details> <summary><b>VS Code (Copilot)</b></summary>
Add to .vscode/mcp.json:
{
"servers": {
"codewiki-mcp": {
"command": "npx",
"args": ["-y", "codewiki-mcp@latest"]
}
}
}
</details>
<details> <summary><b>Local development</b></summary>
{
"mcpServers": {
"codewiki-mcp": {
"command": "node",
"args": ["/path/to/codewiki-mcp/dist/cli.js"]
}
}
}
</details>
💡 Usage
Prompts you can use in any MCP-compatible client:
codewiki fetch how routing works in Next.js
codewiki search state management libraries
codewiki ask how does React fiber reconciler work?
Fetch complete documentation:
codewiki fetch vercel/next.js
codewiki fetch https://github.com/fastify/fastify
Get structured pages:
codewiki fetch pages tailwindlabs/tailwindcss
Ask with natural language:
codewiki ask fastify how to add authentication?
🛠️ MCP Tools
🔍 codewiki_search_repos
Search repositories indexed by codewiki.google.
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
query |
string | ✅ | — | Search query |
limit |
number | — | 10 | Max results (1–50) |
📄 codewiki_fetch_repo
Fetch generated wiki content for a repository.
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
repo |
string | ✅ | — | owner/repo, GitHub URL, or natural-language query |
mode |
string | — | "aggregate" |
"aggregate" — full markdown; "pages" — structured JSON |
💬 codewiki_ask_repo
Ask a natural-language question about a repository.
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
repo |
string | ✅ | — | Repository identifier (same formats as fetch) |
question |
string | ✅ | — | Question about the repo |
history |
array | — | [] |
Conversation history [{role, content}] (max 20) |
📊 Response Format
<details> <summary><b>✅ Success — Search</b></summary>
{
"query": "fastify",
"count": 1,
"items": [
{
"fullName": "fastify/fastify",
"url": "https://github.com/fastify/fastify",
"description": "Fast and low overhead web framework",
"avatarUrl": "https://avatars.githubusercontent.com/u/24939....",
"extraScore": 555
}
],
"meta": {
"totalBytes": 12500,
"totalElapsedMs": 450
}
}
</details>
<details> <summary><b>✅ Success — Fetch (pages mode)</b></summary>
{
"repo": "fastify/fastify",
"commit": "abc123",
"canonicalUrl": "https://github.com/fastify/fastify",
"pages": [
{
"title": "Overview",
"level": 1,
"anchor": "#overview",
"markdown": "# Overview\n\nFastify is a web framework...",
"diagramCount": 1
}
],
"meta": {
"totalBytes": 25000,
"totalElapsedMs": 1200
}
}
</details>
<details> <summary><b>✅ Success — Ask</b></summary>
{
"answer": "Fastify uses a plugin-based architecture where...",
"meta": {
"totalBytes": 8500,
"totalElapsedMs": 2300
}
}
</details>
<details> <summary><b>❌ Error Response</b></summary>
{
"error": {
"code": "RPC_FAIL",
"message": "CodeWiki RPC VSX6ub failed with status 404",
"rpcId": "VSX6ub",
"statusCode": 404
}
}
Error codes: VALIDATION, RPC_FAIL, TIMEOUT, NLP_RESOLVE_FAIL
</details>
⚙️ How It Works
Data Flow
AI Assistant → MCP protocol → codewiki-mcp → HTTPS → codewiki.google
↓
AI Assistant ← MCP protocol ← codewiki-mcp ← JSON ← Google RPC API
Google Batchexecute RPC
codewiki.google uses Google's internal batchexecute RPC format (not REST, not GraphQL). The client:
- Builds a POST request with
f.req=...body - Sends it to
/_/BoqAngularSdlcAgentsUi/data/batchexecute - Receives a response with XSSI prefix
)]}'\n - Parses
wrb.frframes and extracts the typed payload
Each tool maps to an RPC ID:
| Tool | RPC ID |
|---|---|
| 🔍 Search | vyWDAf |
| 📄 Fetch | VSX6ub |
| 💬 Ask | EgIxfe |
🧠 NLP Repo Resolution
Users can type natural language instead of owner/repo:
"the fastify web framework"
→ wink-nlp extracts keyword "fastify" (POS tag: NOUN/PROPN)
→ GitHub Search API: GET /search/repositories?q=fastify&sort=stars
→ top result: "fastify/fastify"
→ normalizeRepoInput("fastify/fastify") → URL for codewiki
🔄 Retry with Exponential Backoff
| Attempt | Delay |
|---|---|
| 0 | immediate |
| 1 | 250ms |
| 2 | 500ms |
| 3 | 1000ms |
4xx errors (client errors) are never retried.
🖥️ CLI
codewiki-mcp [options]
Options:
--http Streamable HTTP transport
--sse SSE transport
--port <number> Port for HTTP/SSE (default: 3000)
--endpoint <str> URL endpoint (default: /mcp)
--help, -h Show help
⚡ Configuration
Environment variables:
| Variable | Default | Description |
|---|---|---|
CODEWIKI_BASE_URL |
https://codewiki.google |
Base URL |
CODEWIKI_REQUEST_TIMEOUT |
30000 |
Request timeout (ms) |
CODEWIKI_MAX_RETRIES |
3 |
Max retries |
CODEWIKI_RETRY_DELAY |
250 |
Base retry delay (ms) |
GITHUB_TOKEN |
— | GitHub token for NLP repo resolution |
You can also create a .env file in the project root:
CODEWIKI_REQUEST_TIMEOUT=60000
CODEWIKI_MAX_RETRIES=5
GITHUB_TOKEN=ghp_your_token
📁 Project Structure
src/
├── cli.ts # CLI entry point
├── server.ts # Transport setup (stdio/HTTP/SSE)
├── index.ts # Library re-exports
├── schemas.ts # Zod input schemas
├── lib/
│ ├── codewikiClient.ts # API client with retry + metadata
│ ├── batchexecute.ts # Google RPC response parser
│ ├── repo.ts # Repo normalization + NLP resolution
│ ├── extractKeyword.ts # NLP keyword extraction (wink-nlp)
│ ├── resolveRepo.ts # GitHub Search API resolver
│ ├── errors.ts # CodeWikiError + formatMcpError
│ └── config.ts # Env-based configuration
└── tools/
├── searchRepos.ts # codewiki_search_repos
├── fetchRepo.ts # codewiki_fetch_repo
└── askRepo.ts # codewiki_ask_repo
❓ Troubleshooting
<details> <summary><b>Permission Denied</b></summary>
chmod +x ./node_modules/.bin/codewiki-mcp
</details>
<details> <summary><b>Connection Refused (HTTP/SSE)</b></summary>
# Check if port is in use
lsof -i :3000
</details>
<details> <summary><b>Timeout Errors</b></summary>
For large repositories, increase the timeout:
CODEWIKI_REQUEST_TIMEOUT=60000 node dist/cli.js
</details>
<details> <summary><b>NLP Resolution Fails</b></summary>
If natural-language input doesn't resolve, use explicit format:
# Instead of "the fastify framework"
fastify/fastify
# or
https://github.com/fastify/fastify
Set GITHUB_TOKEN to avoid GitHub API rate limits for unauthenticated requests.
</details>
🧑💻 Development
npm run dev # stdio with tsx
npm run dev:http # HTTP with tsx
npm run dev:sse # SSE with tsx
npm run typecheck # type check
npm run test # run tests
npm run test:watch # tests in watch mode
npm run build # compile to dist/
🤝 Contributing
Contributions are welcome! Please:
- Fork the repository
- Create a feature branch (
git checkout -b feat/my-feature) - Use Conventional Commits for commit messages
- Run
npm run typecheck && npm run testbefore submitting - Open a Pull Request
📄 License
MIT © codewiki-mcp contributors
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。