notebooklm-mcp

An MCP server that gives Claude (and any MCP-compatible agent) full access to Google NotebookLM — create notebooks, upload sources, ask questions, and generate audio overviews, quizzes, flashcards, mind maps, slide decks, videos, and reports.

Stability note: This server wraps notebooklm-py, which reverse-engineers NotebookLM's unofficial API. It works well for personal and experimental use, but may break without notice if Google changes their API. Not recommended for production/business-critical workflows.

What You Can Do

Related MCP server: notebooklm-mcp-2026

Quick Start

Prerequisites

• Python 3.10+ — check with python3 --version

• uv (recommended) — install with curl -LsSf https://astral.sh/uv/install.sh | sh

• A Google account with NotebookLM access

1. Clone the repo

git clone https://github.com/mulyg/notebooklm-mcp
cd notebooklm-mcp

2. Install dependencies

# With uv (recommended — handles venv automatically)
uv sync

# Or with pip
pip install "notebooklm-py[browser]" "mcp[cli]" pydantic httpx
playwright install chromium

Important: The playwright install chromium step is required for authentication. If you skip it, the server will fail with a browser error.

3. Authenticate with Google

notebooklm login

This opens a browser window to sign in to your Google account. Your session is stored locally — you only need to do this once (sessions last ~30 days).

notebooklm auth check    # verify it worked

What happens under the hood: notebooklm-py uses browser automation (Playwright) to authenticate with Google and store your session cookies in ~/.notebooklm/. No password is saved — only the session token.

Connecting to Agent Environments

Claude Code (CLI)

Add to .mcp.json in your project root, or to ~/.claude/claude_desktop_config.json for global use:

{
  "mcpServers": {
    "notebooklm": {
      "command": "uv",
      "args": ["run", "/path/to/notebooklm-mcp/server.py"]
    }
  }
}

Claude Desktop — macOS

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "notebooklm": {
      "command": "uv",
      "args": ["run", "/Users/yourname/notebooklm-mcp/server.py"]
    }
  }
}

Claude Desktop — Windows

Edit %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "notebooklm": {
      "command": "uv",
      "args": ["run", "C:\\Users\\YourName\\notebooklm-mcp\\server.py"]
    }
  }
}

Note: On Windows, use double backslashes (\\) in JSON paths.

Cursor / Windsurf / other MCP clients

Most MCP-compatible editors use the same command + args format. Point it at server.py using the pattern above. Consult your editor's MCP documentation for the config file location.

Without uv

If you prefer plain Python:

{
  "mcpServers": {
    "notebooklm": {
      "command": "python",
      "args": ["/path/to/notebooklm-mcp/server.py"]
    }
  }
}

Make sure the Python environment that has the dependencies installed is the one being used.

How It Works

Claude / MCP client
       ↓  MCP protocol (stdio)
  server.py  (FastMCP)
       ↓  Python API
  notebooklm-py
       ↓  browser automation (Playwright)
  NotebookLM web API
       ↓
  Google's servers

server.py uses FastMCP to expose 24 tools over the MCP stdio transport. Each tool call passes through notebooklm-py, which uses a headless Chromium browser to talk to NotebookLM's undocumented internal API.

The Generate → Download Pattern

Audio, quiz, flashcards, slides, reports, and videos are generated asynchronously. Calling generate_* starts the generation task. Call download_* to save the result to disk once it's ready.

In practice, ask Claude to do both steps in a single prompt:

"Generate a long deep-dive audio overview of my 'Product Strategy' notebook, then download it to ~/Downloads/strategy.mp3"

Claude will call notebooklm_generate_audio then notebooklm_download_audio automatically.

Mind maps are the exception — notebooklm_generate_mind_map returns the JSON data directly without a separate download step.

Available Tools (24 total)

Start Here

Add Content to a Notebook

Ask Questions

Generate Content (Step 1 of 2)

Download Results (Step 2 of 2)

Example Workflows

Upload & Query

"Create a notebook called 'AI Research', add this arxiv paper URL, then summarise the key findings."

Study Mode

"Generate flashcards and a quiz from my 'Biology 101' notebook. Make the quiz hard. Download both to ~/Downloads as JSON files."

Podcast Creation

"Generate a long deep-dive audio overview of my 'Product Strategy' notebook, then download it to ~/Downloads/strategy-podcast.mp3"

Product Feedback Analysis

"Create a 'Q1 Feedback' notebook. Add these 3 URLs from our support forum. Then ask: what are the top 5 complaints? Finally generate a briefing report."

Research Pipeline

"Research 'gen AI product management trends'. Import the results into my 'PM Research' notebook. Then generate a detailed slide deck from those sources."

With Filesystem MCP (composability)

If you also have the filesystem MCP server active:

"List all PDFs in ~/Documents/reports. Add each one to my 'Annual Reviews' notebook. Then ask: what were the key outcomes across all reports?"

File Upload Guidelines

Supported formats: PDF, DOCX, TXT, audio (MP3, WAV, M4A), video (MP4, MOV), images (PNG, JPG)

Approximate size limits:

• Documents (PDF, DOCX, TXT): up to ~100 MB

• Audio/Video: up to ~500 MB (processing can take 5–10 minutes)

• Images: up to ~50 MB

Tips:

• For large PDFs, consider splitting into sections first

• ~ paths are supported (e.g. ~/Downloads/paper.pdf)

• Video uploads are slower — extracting audio with ffmpeg first can speed things up

Limitations & Caveats

• Unofficial API — notebooklm-py reverse-engineers NotebookLM's private API. Google may change it without notice, breaking the server.

• Experimental only — Not suitable for production or business-critical workflows.

• Rate limiting — NotebookLM has undocumented rate limits. If you hit errors, wait a few minutes and retry.

• Audio generation is slow — Expect 3–7 minutes per audio overview.

• Research polling — The notebooklm_research tool polls for up to 5 minutes. Deep research may occasionally time out; check notebooklm_list_sources afterward to verify.

• No official Google support — This project is not affiliated with or endorsed by Google.

Session Management

Check if your session is active:

notebooklm auth check

Re-authenticate (session expired):

notebooklm login

Reset and start fresh:

rm -rf ~/.notebooklm/
notebooklm login

Clean up test notebooks:

"List my notebooks and delete the ones with 'test' in the name."

Troubleshooting

Authentication failed Your session has expired. Run notebooklm login again.

Browser not installed / Playwright error

playwright install chromium

File not found Use the full absolute path (e.g. /Users/you/docs/paper.pdf or ~/docs/paper.pdf).

Output directory does not exist Make sure the parent directory of your output_path exists before downloading.

mkdir -p ~/Downloads

Audio/video generation times out Audio takes 3–7 minutes; video can take longer. Retry once — the task may still be running in NotebookLM. You can also check https://notebooklm.google.com directly to see if generation completed.

The underlying APIs changed Check for a new release of notebooklm-py and update:

pip install --upgrade notebooklm-py
# or with uv:
uv sync --upgrade

Python version too old

python3 --version   # must be 3.10+

Install Python 3.10+ from python.org or via your package manager.

License

MIT