nblm-mcp

An MCP server that gives an AI agent access to Google NotebookLM (rebranded Gemini Notebook in July 2026): list and create notebooks, manage their sources, ask questions that are answered from those sources with citations, and generate Studio artifacts like audio overviews, briefing docs, quizzes and mind maps.

The point is grounding. NotebookLM answers only from the material you gave it, so an agent that can call ask gets cited answers out of your own documents instead of guessing — and it costs no tokens of your own context, because Gemini does the reading server-side.

⚠️ Unofficial — read this first

Google has no public consumer API for NotebookLM. This server drives the same private web endpoints the notebooklm.google.com UI calls, authenticated with your own browser session cookies, via the MIT-licensed notebooklm-py library.

• Not affiliated with or endorsed by Google.

• The internal API can change without notice and break this server.

• Your Google account's rate limits and daily Studio quotas apply.

• Use an account you are comfortable automating. Best for personal projects, research, and prototypes.

Google does document an official API for Gemini Notebook Enterprise. If you have a Workspace/Cloud org with that feature, prefer it over this.

Install

Not on PyPI yet — install straight from this repository. uvx builds and runs it on demand, so there is nothing to keep updated by hand:

uvx --from git+https://github.com/Diego-Dev-Moros/nblm-mcp nblm-mcp

Log in once

Login needs the auth extra (Playwright) and a human at the keyboard:

uvx --from "nblm-mcp[auth] @ git+https://github.com/Diego-Dev-Moros/nblm-mcp" nblm-mcp-login

A browser window opens; sign in to NotebookLM as you normally would. The session cookies are stored under ~/.notebooklm/ (the same profile layout notebooklm-py uses, so an existing notebooklm login also works). The MCP server never logs in on its own — it needs an interactive browser, which an MCP host cannot provide.

If Playwright has no browser yet, run playwright install chromium first.

Cookies expire. When they do, tools start returning an auth error; run the login command again.

Connect it to a client

Claude Code — -s user makes it available in every project:

claude mcp add notebooklm -s user -- \
  uvx --from git+https://github.com/Diego-Dev-Moros/nblm-mcp nblm-mcp

Claude Desktop — add this to claude_desktop_config.json and restart the app (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\, Linux: ~/.config/Claude/):

{
  "mcpServers": {
    "notebooklm": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Diego-Dev-Moros/nblm-mcp", "nblm-mcp"]
    }
  }
}

Claude Desktop launches from the GUI, which does not inherit your shell's PATH. If the server fails to start there, replace "uvx" with its absolute path (which uvx, typically ~/.local/bin/uvx).

Either way, confirm it works by asking the agent to call auth_status.

Tools

Notes on behavior

• Destructive tools are gated. delete_notebook and delete_source refuse to run without confirm=true, so a stray tool call can't destroy a notebook.

• Generation is slow and quota-bound. generate_artifact returns as soon as the job is queued (wait=false, the default) and tells the agent to poll list_artifacts. Pass wait=true to block instead; it waits up to NBLM_GENERATION_TIMEOUT seconds.

• File paths are server-side. add_source(file_path=...) and download_artifact read and write on the host running the MCP server, which is not necessarily where the user's chat client runs.

Configuration

All optional — see .env.example. A .env in the working directory is loaded if present.

Development

uv venv && uv pip install -e ".[dev]"
uv run pytest
uv run ruff check src tests

The test suite runs the tools against an in-memory fake client — it never touches Google, so it is safe and fast to run anywhere.

Prior art

notebooklm-py does the hard part — reverse-engineering and maintaining the private batchexecute protocol — and ships its own, larger MCP server. This project is a smaller, opinionated tool surface on top of that library: fewer tools, confirmation gates on destructive operations, and citation-resolved answers.

License

MIT — see LICENSE.