granola-mcp

MCP server for semantic search across Granola meeting notes. Extracts insights, themes (pain-points, feature-requests, decisions, etc.), and key quotes with speaker attribution. Uses LanceDB for fast local vector search.

Based on reverse engineering research by Joseph Thacker and getprobo.

Features

• Export: Extract all your Granola meetings with transcripts
• Semantic Search: Vector-indexed search across meetings with pre-extracted insights
• Speaker Attribution: Distinguishes between host (me) and participants
• Theme Extraction: Auto-categorizes content into themes (pain-points, feature-requests, etc.)
• MCP Server: Exposes search to Claude Code, Claude Desktop, and other AI tools

Prerequisites

• Node.js 18+
• Granola desktop app installed and logged in
• OpenAI API key (for embeddings and insight extraction)

Installation

npm install
npm run build

Quick Start

# One command to sync everything (export + index)
OPENAI_API_KEY=sk-... node dist/index.js sync

# Or with a custom data directory
OPENAI_API_KEY=sk-... node dist/index.js sync ./my-data

# Test search
OPENAI_API_KEY=sk-... node dist/index.js search "user pain points"

CLI Commands

Sync (Recommended)

The easiest way to keep your data up to date - exports from Granola and rebuilds the index in one step:

OPENAI_API_KEY=sk-... node dist/index.js sync

# With options
OPENAI_API_KEY=sk-... node dist/index.js sync ./my-data
OPENAI_API_KEY=sk-... node dist/index.js sync --skip-extraction  # Faster, reuses existing insights

Export from Granola

Export only (without indexing):

node dist/index.js export ./output
node dist/index.js export ./output --format markdown
node dist/index.js export ./output --format json

Build Search Index

# Full indexing with insight extraction (~$0.02/document)
OPENAI_API_KEY=sk-... node dist/index.js index ./export

# Skip extraction (use existing insights, just rebuild embeddings)
OPENAI_API_KEY=sk-... node dist/index.js index ./export --skip-extraction

Search from CLI

OPENAI_API_KEY=sk-... node dist/index.js search "pricing concerns"
OPENAI_API_KEY=sk-... node dist/index.js search "feature requests" --folder "User interviews"

Export for ChatGPT

OPENAI_API_KEY=sk-... node dist/index.js export-combined ./chatgpt.md
OPENAI_API_KEY=sk-... node dist/index.js export-combined ./chatgpt.md --query "user feedback"

Other Commands

node dist/index.js list              # List documents
node dist/index.js workspaces        # List workspaces
node dist/index.js folders           # List folders
node dist/index.js transcript <id>   # Get specific transcript

MCP Server Setup

Claude Code

Add to .mcp.json in your project:

{
  "mcpServers": {
    "granola": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/granola-mcp/dist/mcp/server.js"],
      "env": {
        "GRANOLA_DATA_DIR": "/path/to/granola-mcp/export",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "granola": {
      "command": "node",
      "args": ["/path/to/granola-mcp/dist/mcp/server.js"],
      "env": {
        "GRANOLA_DATA_DIR": "/path/to/granola-mcp/export",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

MCP Tools

Speaker Attribution

The system distinguishes between speakers:

• speaker: "me" - The meeting host (you)
• speaker: "participant" - Other people in the meeting

This helps AI understand what's your own idea vs external feedback.

Pre-defined Themes

• pain-points: User frustrations, problems, complaints
• feature-requests: Desired features, wishlist items
• positive-feedback: What users liked, praised
• pricing: Cost concerns, value perception
• competition: Competitor mentions, alternatives
• workflow: How users currently do things
• decisions: Key decisions made, action items
• questions: Open questions needing clarification

Output Structure

export/
├── vectors.lance/           # LanceDB vector index
├── Meeting_Title_1/
│   ├── document.json        # Raw document data
│   ├── notes.md             # Converted notes
│   ├── transcript.json      # Raw transcript with speaker info
│   ├── transcript.md        # Formatted transcript
│   └── transcript.txt       # Plain text transcript
└── Meeting_Title_2/
    └── ...

Keeping Data Updated

The system doesn't auto-sync with Granola. Run sync manually after new meetings, or set up a cron job:

Manual Update

OPENAI_API_KEY=sk-... node dist/index.js sync

Automated Updates (Cron)

Add to your crontab (crontab -e):

# Sync every night at 2am
0 2 * * * cd /path/to/granola-mcp && OPENAI_API_KEY=sk-... /usr/local/bin/node dist/index.js sync >> /tmp/granola-sync.log 2>&1

# Or every 6 hours
0 */6 * * * cd /path/to/granola-mcp && OPENAI_API_KEY=sk-... /usr/local/bin/node dist/index.js sync >> /tmp/granola-sync.log 2>&1

macOS LaunchAgent

Create ~/Library/LaunchAgents/com.granola-mcp.sync.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.granola-mcp.sync</string>
    <key>ProgramArguments</key>
    <array>
        <string>/usr/local/bin/node</string>
        <string>/path/to/granola-mcp/dist/index.js</string>
        <string>sync</string>
    </array>
    <key>EnvironmentVariables</key>
    <dict>
        <key>OPENAI_API_KEY</key>
        <string>sk-...</string>
    </dict>
    <key>StartCalendarInterval</key>
    <dict>
        <key>Hour</key>
        <integer>2</integer>
        <key>Minute</key>
        <integer>0</integer>
    </dict>
    <key>StandardOutPath</key>
    <string>/tmp/granola-sync.log</string>
    <key>StandardErrorPath</key>
    <string>/tmp/granola-sync.log</string>
</dict>
</plist>

Load it with: launchctl load ~/Library/LaunchAgents/com.granola-mcp.sync.plist

How It Works

1. Export: Reads credentials from ~/Library/Application Support/Granola/supabase.json and fetches all documents via Granola's API

2. Index:

   • Extracts themes and key quotes using GPT-4o-mini
   • Generates embeddings using text-embedding-3-small
   • Stores in LanceDB for fast vector search

3. Search:

   • Embeds your query
   • Finds semantically similar documents
   • Returns summaries + relevant quotes (not raw transcripts)

Cost Estimates

Search queries are free (vector similarity, no LLM calls).

License

MIT