MCP Documentation Server

A TypeScript-based Model Context Protocol (MCP) server that provides document management and semantic search capabilities. Upload documents, search them with AI embeddings, and integrate seamlessly with MCP clients like Claude Desktop.

Demo Video

Quick Start

1. Install and Run

# Run directly with npx (recommended)
npx @andrea9293/mcp-documentation-server

2. Configure MCP Client

Add to your MCP client configuration (e.g., Claude Desktop):

{
  "mcpServers": {
    "documentation": {
      "command": "npx",
      "args": [
        "-y",
        "@andrea9293/mcp-documentation-server"
      ],
      "env": {
        "MCP_EMBEDDING_MODEL": "Xenova/all-MiniLM-L6-v2"
      }
    }
  }
}

3. Start Using

• Add documents: Upload text/markdown files or add content directly
• Search documents: Use semantic search to find relevant information
• Manage content: List, retrieve, and organize your documents

Features

• 📄 Document Management - Add, list, retrieve, and delete documents with metadata
• 🔍 Semantic Search - AI-powered search using embeddings
• 🧠 Intelligent Chunking - Documents are automatically split into context-aware chunks for better search accuracy and context retrieval
• 🧩 Context Window Retrieval - Retrieve a window of chunks around a relevant chunk to get more context for your queries
• � LLM Guidance - The search tool provides hints to LLMs on how to use context windowing for better answers
• 📁 File Upload - Drop .txt/.md/.pdf files in uploads folder for processing
• 🗑️ Document Deletion - Clean removal of documents and their chunks
• 🌍 Multilingual - Supports multiple languages with quality embeddings
• 💾 Local Storage - All data stored locally in ~/.mcp-documentation-server/ directory
• ⚡ Fast Setup - No database required, works out of the box

Available Tools

Usage Examples

Adding a Document

{
  "tool": "add_document",
  "arguments": {
    "title": "Python Basics",
    "content": "Python is a high-level programming language...",
    "metadata": {
      "category": "programming",
      "tags": ["python", "tutorial"]
    }
  }
}

Searching Documents

{
  "tool": "search_documents",
  "arguments": {
    "document_id": "doc-123",
    "query": "variable assignment",
    "limit": 5
  }
}

Retrieving Context Window

{
  "tool": "get_context_window",
  "arguments": {
    "document_id": "doc-123",
    "chunk_index": 5,
    "before": 2,
    "after": 2
  }
}

Deleting a Document

{
  "tool": "delete_document",
  "arguments": {
    "id": "doc-123"
  }
}

File Upload Workflow

1. Get uploads path: get_uploads_path (~/.mcp-documentation-server/uploads/)
2. Place your .txt/.md/.pdf files in that folder
3. Process files: process_uploads
4. Search the processed documents

Supported file types:

• .txt - Plain text files
• .md - Markdown files
• .pdf - PDF files (text extraction, no OCR)

Configuration

Data Storage

All documents and uploads are stored locally in:

~/.mcp-documentation-server/
├── data/      # Document storage (JSON files)
└── uploads/   # Files to process (.txt, .md, .pdf)

Embedding Models

Set via MCP_EMBEDDING_MODEL environment variable:

• Xenova/all-MiniLM-L6-v2 (default) - Fast, good quality (384 dimensions)
• Xenova/paraphrase-multilingual-mpnet-base-v2 (recommended) - Best quality, multilingual (768 dimensions)

The system automatically manages the correct embedding dimension for each model. Embedding providers expose their dimension via getDimensions().

⚠️ Important: Changing models requires re-adding all documents as embeddings are incompatible.

Installation Options

NPX (Recommended)

npx @andrea9293/mcp-documentation-server

Global Installation

npm install -g @andrea9293/mcp-documentation-server
mcp-documentation-server

From Source

git clone https://github.com/andrea9293/mcp-documentation-server.git
cd mcp-documentation-server
npm install
npm run build
npm start

Best Practices

Document Organization

• Use descriptive titles for easy identification
• Add relevant metadata (tags, categories) for better organization
• Keep documents focused on specific topics for better search accuracy

Search Optimization

• Use specific, descriptive search queries
• Combine keywords related to your topic
• Start with broader queries, then refine with more specific terms
• After finding relevant chunks with search_documents, use get_context_window to retrieve additional context around those chunks. You can call get_context_window multiple times until you have enough context to answer your question.

Performance Tips

• Process large files during off-peak hours (initial embedding creation)
• Use smaller embedding models for faster performance if quality is acceptable
• Regularly clean up unused documents to maintain performance

Troubleshooting

Timeout on First Use

• Cause: Embedding models download on first use (~420MB for best model)
• Solution: Wait for background download to complete, or use smaller model initially

Search Results Issues

• Cause: Mixed embedding models in same dataset
• Solution: Stick to one model or re-add all documents after switching

Development

# Development server with hot reload
npm run dev

# Build and test
npm run build

# Inspect tools with web UI
npm run inspect

Contributing

1. Fork the repository
2. Create feature branch: git checkout -b feature/name
3. Follow Conventional Commits for messages
4. Submit pull request

License

MIT - see LICENSE file

Support

• 📖 Documentation
• 🐛 Report Issues
• 💬 MCP Community

Built with FastMCP and TypeScript 🚀

Star History