Bear Notes MCP Server with RAG

Looking to supercharge your Bear Notes experience with AI assistants? This little gem connects your personal knowledge base to AI systems using semantic search and RAG (Retrieval-Augmented Generation).

I built this because I wanted my AI assistants to actually understand what's in my notes, not just perform simple text matching. The result is rather sweet, if I do say so myself.

Getting Started

Setting up is straightforward:

git clone [your-repo-url]
cd bear-mcp-server
npm install

Make the scripts executable (because permissions matter):

chmod +x src/bear-mcp-server.js
chmod +x src/create-index.js

First Things First: Index Your Notes

Before diving in, you'll need to create vector embeddings of your notes:

npm run index

Fair warning: this might take a few minutes if you're a prolific note-taker like me. It's converting all your notes into mathematical vectors that capture their meaning— clever stuff 😉.

Configuration

Update your MCP configuration file:

{
  "mcpServers": {
    "bear-notes": {
      "command": "node",
      "args": [
        "/absolute/path/to/bear-mcp-server/src/bear-mcp-server.js"
      ],
      "env": {
        "BEAR_DATABASE_PATH": "/Users/yourusername/Library/Group Containers/9K33E3U3T4.net.shinyfrog.net.bear/Application Data/database.sqlite"
      }
    }
  }
}

🚨 Remember to replace the path with your actual installation location. No prizes for using the example path verbatim, I'm afraid.

What Makes This Special?

Semantic Search: Find notes based on meaning, not just keywords. Ask about "productivity systems" and it'll find your notes on GTD and Pomodoro, even if they don't contain those exact words.
RAG Support: Your AI assistants can now pull in relevant context from your notes, even when you haven't explicitly mentioned them.
All Local Processing: Everything runs on your machine. No data leaves your computer, no API keys needed, no internet dependency (after initial setup).
Graceful Fallbacks: If semantic search isn't available for whatever reason, it'll quietly fall back to traditional search. Belt and braces.

How It Works

The Clever Bits

This server uses the Xenova implementation of transformers.js with the all-MiniLM-L6-v2 model:

It creates 384-dimensional vectors that capture the semantic essence of your notes
All processing happens locally on your machine
The first startup might be a tad slow while the model loads, but it's zippy after that

The Flow

Your query gets converted into a vector using the transformer model
This vector is compared to the pre-indexed vectors of your notes
Notes with similar meanings are returned, regardless of exact keyword matches
AI assistants use these relevant notes as context for their responses

Project Structure

Nothing too complex here:

bear-mcp-server/
├── package.json
├── readme.md
└── src/
    ├── bear-mcp-server.js     # Main MCP server
    ├── create-index.js        # Script to index notes
    ├── utils.js               # Utility functions
    ├── lib/                   # Additional utilities and diagnostic scripts
    │   └── explore-database.js # Database exploration and diagnostic tool
    ├── note_vectors.index     # Generated vector index (after indexing)
    └── note_vectors.json      # Note ID mapping (after indexing)

Available Tools for AI Assistants

AI assistants connecting to this server can use these tools:

search_notes: Find notes that match a query
- Parameters: query (required), limit (optional, default: 10), semantic (optional, default: true)
get_note: Fetch a specific note by its ID
- Parameters: id (required)
get_tags: List all tags used in your Bear Notes
retrieve_for_rag: Get notes semantically similar to a query, specifically formatted for RAG
- Parameters: query (required), limit (optional, default: 5)

Requirements

Node.js version 16 or higher
Bear Notes for macOS
An MCP-compatible AI assistant client

Limitations & Caveats

Read-only access to Bear Notes (we're not modifying your precious notes)
macOS only (sorry Windows and Linux folks)
If you add loads of new notes, you'll want to rebuild the index with npm run index
First startup is a bit like waiting for the kettle to boil while the embedding model loads

Troubleshooting

If things go wonky:

Double-check your Bear database path
Make sure you've run the indexing process with npm run index
Check permissions on the Bear Notes database
Verify the server scripts are executable
Look for error messages in the logs

When in doubt, try turning it off and on again. Works more often than we'd like to admit.

License

MIT (Feel free to tinker, share, and improve)

Integrations