mcp-local-rag

Local MCP server for RAG over PDFs, DOCX, images, and plaintext files.

Requirements

For more complex PDFs and image files, the following environment variables can be provided:

• AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT; requires mcp-local-rag[azure].
• AZURE_DOCUMENT_INTELLIGENCE_KEY; when omitted, DefaultAzureCredential is used. Requires mcp-local-rag[azure].
• GEMINI_API_KEY
• MCP_LOCAL_RAG_GEMINI_MODEL (default: gemini-3-pro-preview)

Image files require either Gemini or Azure DI for extraction. See docs/configuration.md for details.

Data Storage

By default, the server stores data in:

• Windows: %LOCALAPPDATA%\mcp-local-rag\
• macOS: ~/Library/Application Support/mcp-local-rag/
• Linux: $XDG_DATA_HOME/mcp-local-rag/

The data directory contains:

• markdown/ - Extracted Markdown content of indexed documents
• metadata.db - SQLite database for document/collection metadata
• qdrant/ - Vector database for embeddings

AI Models are cached in the default HuggingFace cache directory (~/.cache/huggingface/).

To customize the data directory, set the MCP_LOCAL_RAG_DATA_DIR environment variable (a mcp-local-rag/ subfolder is created automatically inside it).

Usage

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "mcp-local-rag": {
      "command": "uvx",
      "args": [
        "--python",
        "3.13",  // Does not support Python 3.14 yet: https://github.com/microsoft/markitdown/issues/1470
        "mcp-local-rag@latest"
      ]
    }
  }
}

If you run into SSL errors (Zscaler), you can try:

{
  "servers": {
    "mcp-local-rag": {
      "command": "uvx",
      "args": [
        "--native-tls",
        "--python",
        "3.13",  // Does not support Python 3.14 yet: https://github.com/microsoft/markitdown/issues/1470
        "--with",
        "pip-system-certs",
        "mcp-local-rag@latest"
      ]
    }
  }
}