MyDocsMCP: MCP Server for PDF Collections

This project is a Model Context Protocol (MCP) Server that enables semantic search (local RAG) over a collection of PDF documents. It uses the FastMCP framework, the ChromaDB vector database, and local embedding models from Sentence Transformers.

Architecture

• Semantic Search: 100% local (offline) RAG (Retrieval-Augmented Generation).

• Embeddings: paraphrase-multilingual-mpnet-base-v2 (supports Portuguese).

• Vector DB: Persistent ChromaDB.

• Watcher: Monitors new PDFs in the ./data/pdfs folder and indexes them automatically via watchdog.

Related MCP server: DocScan MCP Server

How to Use

1. Data Preparation

Place your PDFs in the ./data/pdfs/ folder. If you want to organize them by disciplines, create subfolders:

data/pdfs/
  ├── Generative-AI/
  │   └── lecture1.pdf
  └── Machine-Learning/
      └── fundamentals.pdf

The subfolder name will be used as the discipline metadata.

2. Extremely Simple Configuration (Claude / Gemini Desktop)

To use the server, add the configuration below to your agent's JSON file (claude_desktop_config.json or Gemini's settings.json).

Claude Path (macOS): ~/Library/Application Support/Claude/claude_desktop_config.json Gemini Path (macOS): ~/.gemini/settings.json

The server automatically resolves all data folders (pdfs, metadata, chroma_db) based on the project root. You only need to provide the absolute path where you cloned the repository:

{
  "mcpServers": {
    "mydocsmcp": {
      "command": "uv",
      "args": [
        "--directory", "/Absolute/Path/To/Your/MyDocsMCP",
        "run",
        "mydocs-mcp"
      ]
    }
  }
}

That's it! No additional environment variables (PYTHONPATH, PDF_DIR, etc.) are required. The setup "Just Works"™.

Exposed Tools

• search_documents(query, top_k=5, discipline=None): Semantic search in the collection.

• list_documents(discipline=None): Lists indexed PDFs.

• cross_topic_search(query, disciplines): Cross-topic search across multiple disciplines.

• get_index_stats(): Vector database statistics.

• ingest_new_documents(path=None, force_reindex=False): Forces manual re-ingestion.

Local Development (Python)

We use the uv package manager:

# Install dependencies
uv sync

# Run the server
uv run mydocs-mcp

Running Tests

uv run pytest

Technologies Used

• FastMCP

• uv

• ChromaDB

• PyMuPDF

• Sentence-Transformers

• Watchdog