The official server
Google publishes a remote MCP server for Gemini Notebook, the name NotebookLM has carried since July 2026. It lives at https://notebook.googleapis.com/mcp, speaks Streamable HTTP, and is part of the Gemini Notebook API, which Google calls an early access Public Preview.
It has seven tools:
| Tool group | What it does |
|---|---|
| Notebooks | Create, get, list (with an ownership filter), update the display name |
| Sources | Create (a web URL, a YouTube video, plain text or Markdown), get (with status PROCESSING, ACTIVE or FAILED), list |
So an agent can build and fill a notebook: “make a notebook for this tender and add these ten URLs”. What the tool list does not include is asking the notebook something, or generating an Audio Overview. You still do that in Gemini Notebook itself.
How to connect it
- Use a personal Google account that has opened Gemini Notebook at least once. Workspace accounts are not supported yet.
- Enable the Gemini Notebook API in a Google Cloud project.
- Log in with Application Default Credentials and request the scope
https://www.googleapis.com/auth/notebook. A plaingcloud auth print-access-tokendoes not carry that scope. - Mint an access token and send it in the
Authorizationheader, with your project inx-goog-user-project. - Claude Code: add it with
claude mcp add gemini-notebook-api --transport http, the server URL and both headers. Cursor: add it undermcpServersin.cursor/mcp.jsonin the project, or in~/.cursor/mcp.jsonto use it everywhere. - Refresh the token when it expires; Google says tokens minted this way last one hour.
The rate limits of the API apply: per user 20 new sources a minute and 60 calls a minute for every other method, and the plan of the account still caps sources per notebook (50 free, up to 600 on Ultra). The API is free during the preview.
The unofficial servers
Search for a NotebookLM MCP server and most results are community projects. They can do more than Google’s server, including asking a notebook questions and generating outputs, because they drive the same internal endpoints as the web app. notebooklm-py, for example, ships an MCP server next to its Python library, and its README says plainly: “Unofficial Library - Use at Your Own Risk”, built on undocumented Google APIs that can change without notice.
For a company, three risks come with that:
- The login is you. These servers log in through a browser or reuse your browser cookies. notebooklm-py’s optional master token is, in its own words, a powerful full-account credential.
- It can break any day. Google can change the internal endpoints without notice.
- Nobody signed for the data. Documents go through a tool that is not covered by your agreement with Google. Our checklist is in Is MCP safe to use in a company?, and how to judge a server is in Where to find trustworthy MCP servers.
The common mistake: expecting Claude to read the notebook
People connect NotebookLM to Claude Code hoping Claude will now “know” what is in their notebooks. With the official server, the flow runs the other way: Claude can feed sources into a notebook, not pull answers out. With an unofficial server Claude can ask questions, but each answer is a fresh lookup in one notebook; nothing of it stays with Claude for the next session, and nothing reaches a colleague’s Cursor or ChatGPT.
What this means for a team
MCP connects one agent to one notebook, under one person’s Google account. That is useful for a researcher who lives in Claude Code and keeps sources in Gemini Notebook. It does not turn the notebooks into team knowledge: each developer sets up their own connection and token, Workspace accounts are out for now, and every notebook stays separate. For the bigger picture of MCP in a team, see What is an MCP memory server? and the programmatic side in Does NotebookLM have an API?
Where Bombos fits
Bombos is a plugin that connects to Claude, Claude Code, Cursor, ChatGPT, Copilot and your own agents and gives them all one shared memory: team knowledge, rules it learns from corrections, and what already sits in mail, documents and repos, with a trail of what each answer rests on. It runs in the cloud or on your own server. One developer can try it first with their own agents, next to the notebooks they already keep.
Related questions
Does the official NotebookLM MCP server work with a Workspace account?
Not yet. Google's page states that Google Workspace accounts are not yet supported; the preview is for personal Google accounts.
Can Claude ask my notebook a question through the official MCP server?
Not with the documented tools. The seven tools create, get, list and rename notebooks and create, get and list sources. Asking questions is not among them.
Why does my connection stop working after an hour?
Google notes that access tokens minted with gcloud auth application-default print-access-token expire after one hour. Mint a new one and update the header.
Is an unofficial NotebookLM MCP server safe?
It works, but it logs in as you with browser cookies or a token that can act on your whole Google account, and it relies on endpoints Google does not document. Use a separate account and keep company documents out of it.
