DeepWiki MCP Server
Enables Claude to explore, analyze, and answer questions about any public GitHub codebase by connecting to DeepWiki's AI-powered documentation.
README
DeepWiki MCP Server
Give Claude superpowers to understand any codebase instantly.
This MCP (Model Context Protocol) server connects Claude to DeepWiki, an AI-powered platform by Cognition Labs (creators of Devin AI) that provides "Deep Research for GitHub" - interactive, up-to-date documentation for any public repository. With this integration, Claude can explore, analyze, and answer questions about any GitHub codebase - without you having to manually copy-paste code or documentation.
⚠️ Disclaimer
This is a learning/experimental project created through AI-assisted "vibe coding".
- The author is not a professional developer
- This entire MCP server was built using Claude Opus 4.0 (vibe coding)
- Code quality improvements and auditing done with Claude Sonnet 4.5
- Provided as-is for educational and experimental purposes only
- Use at your own risk - see LICENSE for full disclaimer
What is DeepWiki?
DeepWiki is an AI platform by Cognition Labs (the team behind Devin AI) that automatically transforms any public GitHub repository into interactive, conversational documentation.
How it works:
- Visit any repo:
github.com/facebook/react - Change URL to:
deepwiki.org/facebook/react - Get instant access to AI-generated architecture diagrams, documentation, and an interactive chatbot
DeepWiki analyzes the code structure, relationships, and patterns, then creates up-to-date documentation you can have conversations with - like having an expert who has deeply studied the codebase and can answer questions about it.
What Does This MCP Do?
This MCP server acts as a command-line tool that Claude can use automatically when you ask questions about code. When you chat with Claude Desktop and mention a GitHub repository, Claude can:
- Read Documentation: Fetch and parse DeepWiki's generated documentation for any repo
- Ask Questions: Query DeepWiki's AI to understand how code works, find implementations, or explore architecture
- Get Code Snippets: Retrieve exact code references that answer your questions
- Deep Research: Trigger comprehensive analysis for complex questions (3-15 minutes)
- Cache Results: Store responses locally for instant retrieval in follow-up questions
You don't directly use this tool - instead, you talk to Claude naturally, and Claude decides when to use these tools to help answer your questions.
Key Benefits
✅ Understand unfamiliar codebases instantly - No need to clone repos or wade through docs ✅ Ask natural questions - "How does authentication work?" instead of reading thousands of lines ✅ Get exact code references - Automatically retrieves relevant snippets with file paths and line numbers ✅ Follow-up conversations - Ask deeper questions based on previous answers ✅ Save research - Export findings to markdown files for documentation ✅ Works with any public GitHub repo - React, Next.js, your company's repos, etc.
Installation
Prerequisites
- Node.js 18 or higher
- Claude Desktop app
Setup Steps
-
Clone this repository:
git clone https://github.com/ai-vivid/deepwiki-mcp.git cd deepwiki-mcp -
Install dependencies:
npm install -
Install Playwright browsers (required for web automation):
npx playwright install chromium -
Build the project:
npm run build -
Add to Claude Desktop configuration:
On macOS:
~/Library/Application Support/Claude/claude_desktop_config.jsonOn Windows:%APPDATA%\Claude\claude_desktop_config.json{ "mcpServers": { "deepwiki": { "command": "node", "args": ["/absolute/path/to/deepwiki-mcp/dist/index.js"] } } }Replace
/absolute/path/to/deepwiki-mcpwith your actual installation path. -
Restart Claude Desktop
How to Use
Just chat naturally with Claude! Here are example conversations:
Example 1: Understanding a new framework
You: "Can you explain how Next.js handles server-side rendering?"
Claude: *uses wiki_question tool to query DeepWiki about vercel/next.js*
Claude: "Next.js handles SSR through..."
Example 2: Finding implementations
You: "Show me how React implements hooks"
Claude: *retrieves code snippets from facebook/react*
Claude: "Here's the hooks implementation from ReactFiberHooks.js..."
Example 3: Deep research
You: "I need a comprehensive explanation of how Kubernetes manages container orchestration"
Claude: *triggers deep research mode for kubernetes/kubernetes*
Claude: "Starting deep research - this will take 5-10 minutes..."
*returns detailed architectural analysis*
Available Tools (For Claude's Use)
Claude automatically decides when to use these tools. This section explains what each tool does for your understanding, but you don't need to call them directly.
Tool 1: wiki_parser
Purpose: Reads DeepWiki's static documentation pages for a repository.
When Claude uses it:
- You ask for an overview of a project's structure
- You want to see the table of contents for documentation
- You request specific chapters from the documentation
Parameters Claude can set:
| Parameter | What it does | Example value |
|---|---|---|
repo |
GitHub repository to query | "facebook/react" |
action |
"structure" = show table of contents<br>"extract" = get specific content |
"extract" |
chapters |
Specific chapters to retrieve | ["Getting Started", "API Reference"] |
depth |
How many heading levels to show (1-4) | 2 |
saveToFile |
"save-only" = just save to file<br>"save-and-show" = save and display |
"save-and-show" |
Example - What Claude does:
// When you ask: "Show me the React documentation structure"
wiki_parser({
repo: "facebook/react",
action: "structure",
depth: 2
})
// Returns: A table of contents with 2 levels of headers
// Output looks like:
// 1: Getting Started
// ## Installation
// ## Quick Start
// 2: Main Concepts
// ## Components
// ## Props & State
Example - Extracting specific content:
// When you ask: "Get me the Getting Started guide from React docs"
wiki_parser({
repo: "facebook/react",
action: "extract",
chapters: ["Getting Started"]
})
// Returns: Full markdown content of the "Getting Started" chapter
Tool 2: wiki_question
Purpose: Asks DeepWiki's AI questions about a repository and gets intelligent answers with code references.
When Claude uses it:
- You ask how something works in a codebase
- You want to find specific implementations
- You need architectural explanations
- You ask follow-up questions about previous answers
Parameters Claude can set:
| Parameter | What it does | Example value |
|---|---|---|
repo |
GitHub repository to query | "facebook/react" |
question |
Your question (for new queries) | "How does React implement hooks?" |
queryId |
ID from previous answer (for follow-ups) | "query-abc123" |
useDeepResearch |
Enable 3-15 minute deep analysis | true |
goDeeper |
Get more details on previous query | true |
followUpQuestion |
Ask follow-up on same topic | "What about error handling?" |
includeAnswer |
Show the text explanation | true (default) |
includeReferencesList |
Show numbered file references | true (default) |
referencesNumbers |
Get specific code snippets by number | [1, 2, 3] |
contextFiles |
Get full content of specific files | ["src/hooks.js"] |
contextRanges |
Get specific line ranges | {"src/hooks.js": {"start": 1, "end": 50}} |
saveToFile |
Save response to file | "save-and-show" |
Example 1 - Basic question:
// When you ask: "How does React handle component state?"
wiki_question({
repo: "facebook/react",
question: "How does React handle component state?"
})
// Returns:
// Query ID: query-abc123
//
// # Answer
// React handles component state through the useState hook...
// [1] [2]
//
// # References
// [1]: facebook/react: src/ReactHooks.js:45-67
// [2]: facebook/react: src/ReactFiberHooks.js:120-145
Example 2 - Getting specific code snippets:
// When you say: "Show me the code for reference 1 from that last answer"
wiki_question({
repo: "facebook/react",
queryId: "query-abc123", // From previous response
includeAnswer: false, // Don't repeat the explanation
includeReferencesList: false, // Don't show the reference list again
referencesNumbers: [1] // Just show code for reference #1
})
// Returns:
// Query ID: query-abc123
//
// # Referenced Files
// ## facebook/react: src/ReactHooks.js
// **[45-67]:**
// ```
// function useState(initialState) {
// const hook = mountState(initialState);
// return [hook.state, hook.dispatch];
// }
// ...
// ```
Example 3 - Deep research mode:
// When you ask: "Give me a comprehensive analysis of Next.js routing architecture"
wiki_question({
repo: "vercel/next.js",
question: "Explain the complete routing architecture",
useDeepResearch: true,
saveToFile: "save-and-show"
})
// What happens:
// 1. DeepWiki spends 3-15 minutes doing deep analysis
// 2. Returns comprehensive architectural breakdown
// 3. Saves to: ~/.deepwiki-mcp/output/vercel-next.js/questions/2025-09-30_14-30-00_routing-architecture_query-xyz789.md
// 4. Shows you the full analysis
Example 4 - Follow-up questions:
// You: "What about error handling in that routing system?"
wiki_question({
repo: "vercel/next.js",
queryId: "query-xyz789", // References previous deep research
followUpQuestion: "How does error handling work in the routing system?",
includeFullConversation: false // Only show new answer, not previous one
})
// Returns: Answer about error handling, building on previous context
Example 5 - Getting full file contents:
// You: "Show me the full content of that routing file"
wiki_question({
repo: "vercel/next.js",
queryId: "query-xyz789",
includeAnswer: false,
contextFiles: ["packages/next/src/server/router.ts"]
})
// Returns: Complete contents of router.ts file
Understanding API Polling (Advanced)
When you ask DeepWiki a question, it doesn't respond instantly. Instead, your question is processed asynchronously by their AI, and this MCP periodically checks if the answer is ready.
How it works:
- Question submitted → DeepWiki starts processing
- MCP checks for answer → Polls every few seconds/minutes
- Answer ready → Returns results to Claude
Polling schedules:
-
Regular mode: Checks at 10s, 15s, 20s, 30s, 45s, 75s, 2m, 3m, 5m
- Total: up to 5 minutes for an answer
-
Deep research mode: Checks at 3m, 4m, 5m, 7m, 9m, 12m, 15m
- Total: up to 15 minutes for comprehensive analysis
Why this matters:
- Regular questions: Usually answered in 30-60 seconds
- Deep research: Takes 5-15 minutes but provides much more thorough analysis
- You can customize these intervals (see Configuration below)
Visual example:
You ask question → DeepWiki AI thinks → MCP checks → Gets answer → Claude shows you
Regular: [0s] ----10s----15s----20s---[Answer!]
Deep: [0s] ----------------3m--------------5m---------7m----[Answer!]
Configuration
You can customize the MCP's behavior through environment variables in your Claude Desktop config.
File Storage
default_directory - Where to save exported files
- Default:
~/.deepwiki-mcp/output - Example:
"/Users/me/Documents/deepwiki-research"
allowed_directories - Security: which directories can be written to
- Default:
~/.deepwiki-mcpanddefault_directory - Example:
"/Users/me/Documents/deepwiki-research,/Users/me/projects"
API Polling
DEEPWIKI_POLL_INTERVALS_REGULAR - When to check for regular answers (in seconds)
- Default:
"10,15,20,30,45,75,120,180,300"(checks at these intervals) - Faster polling:
"5,10,15,20,30,60"(checks more frequently)
DEEPWIKI_POLL_INTERVALS_DEEP - When to check for deep research answers (in seconds)
- Default:
"180,240,300,420,540,720,900"(3min, 4min, 5min, 7min, 9min, 12min, 15min) - Faster polling:
"60,120,180,300,600"(checks more frequently)
Example Configuration
{
"mcpServers": {
"deepwiki": {
"command": "node",
"args": ["/path/to/deepwiki-mcp/dist/index.js"],
"env": {
"default_directory": "/Users/me/Documents/deepwiki-research",
"allowed_directories": "/Users/me/Documents/deepwiki-research,/Users/me/projects",
"DEEPWIKI_POLL_INTERVALS_REGULAR": "5,10,15,30,60",
"DEEPWIKI_POLL_INTERVALS_DEEP": "60,120,240,480"
}
}
}
}
Caching & Storage
Cache location: ~/.deepwiki-mcp/cache/
- Stores previous responses for instant retrieval
- Organized by repository
- Automatically used for follow-up questions
Output location: ~/.deepwiki-mcp/output/ (or your default_directory)
- Saved files use this structure:
{repo}/{type}/{date}_{time}_{description}_query-{id}.md - Example:
facebook-react/questions/2025-09-30_14-30-00_hooks-implementation_query-abc123.md
Tips & Best Practices
💡 Start with regular questions - Use deep research only for complex architectural questions 💡 Use follow-ups - Build on previous answers instead of asking everything at once 💡 Save important findings - Ask Claude to save responses for future reference 💡 Be specific - "How does React implement the useState hook?" vs "How does React work?" 💡 Request code when needed - "Show me the code for reference 2" to get actual implementation
Troubleshooting
"No cached response found" → The queryId you provided doesn't exist. Ask a new question first.
"Tool timeout" → DeepWiki is taking longer than expected. Try again or increase polling intervals.
"Repository not found"
→ Check the repo format is owner/repo and that it's a public GitHub repository.
Playwright errors
→ Run npx playwright install chromium to install browser dependencies.
Development
Project Structure
deepwiki-mcp/
├── src/
│ ├── index.ts # MCP server entry point
│ ├── tools/ # Tool implementations
│ │ ├── wiki-parser.ts # Documentation parser tool
│ │ └── wiki-question.ts # Question/answer tool
│ ├── automation/ # Playwright browser automation
│ ├── parsers/ # Response parsers & transformers
│ ├── cache/ # Cache management
│ └── utils/ # Helper utilities
├── dist/ # Compiled JavaScript (generated)
├── package.json
└── tsconfig.json
Building
npm run build
Development Mode
npm run dev
License
MIT - See LICENSE file for full details including disclaimer.
Acknowledgments
- Built with Model Context Protocol
- Powered by DeepWiki (by Cognition Labs)
- Created using Claude (Opus 4.0 & Sonnet 4.5)
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。