Gemini Notebook MCP servers compared

Search for a Gemini Notebook MCP server — the product called NotebookLM until Google renamed it in July 2026 — and you find at least five projects, four of them alive, plus a layer of directory pages ranking them without measuring anything.

So here are the measurements, taken the same afternoon, with the commands to retake them. This project is one of the five, and it is not the biggest: that is stated below rather than buried.

The numbers, 7 November 2026

Project	Stars	Language	Installs / 30 days	Last push
teng-lin/notebooklm-py	20 151	Python	134 206 (PyPI notebooklm-py)	same day
jacob-bd/gemini-notebook-mcp-cli	6 472	Python	66 615 (notebooklm-mcp-cli + notebooklm-mcp-server)	day before
PleasePrompto/notebooklm-mcp	3 443	TypeScript	11 724 (npm notebooklm-mcp)	archived 10 Sep 2026
roomi-fields/notebooklm-mcp (this one)	228	TypeScript	5 411 (npm @roomi-fields/notebooklm-mcp)	same day
PavelGuzenfeld/notebooklm-mcp	2	Python	45 (PyPI mcp-notebooklm)	Sep 2026

Retake it yourself — no key, no account:

gh api repos/teng-lin/notebooklm-py --jq '.stargazers_count'
curl -s https://pypistats.org/api/packages/notebooklm-py/recent | jq .data.last_month
curl -s https://api.npmjs.org/downloads/point/last-month/notebooklm-mcp | jq .downloads

Read install counts as orders of magnitude, not headcounts: mirrors, proxies and continuous-integration runs are in there, and a fresh package can show dozens of downloads within hours of a release with nobody involved.

What each one is shaped for

notebooklm-py is a Python library first, with a CLI, an MCP server and an agent skill around it. It covers the internal API more broadly than anything else here, and it is used thirty times more than the next TypeScript project. If you write Python, start there — we read its source before building anything new, and say so in our own contributor notes.

gemini-notebook-mcp-cli is a command line with an MCP server attached, 43 tools by its own count, and the only project that put the current product name in its repository name — which is why it owns the gemini notebook mcp search outright.

PleasePrompto/notebooklm-mcp was archived on 10 September 2026 and is still the second-most installed package in this list, which is worth knowing before you build on it: it works until Google next changes something. This project grew out of it — 17 of its 19 tool names still work here unchanged — and the migration page exists for exactly that.

This project is the only one where the HTTP REST API is the product rather than a side effect: 56 endpoints, the same 56 as MCP tools, the same 56 described in a UTCP manual, and typed clients in Rust and Ruby generated from the specification. That shape exists because an agent is not the only caller: n8n, Zapier, Make, cron and curl all speak HTTP and none of them speak MCP.

mcp-notebooklm wraps notebooklm-py for MCP. At 45 installs a month it is a weekend project, listed here because directories list it.

Where this one loses

• Reach. 228 stars against 20 151, and 5 411 installs against 134 206. If you are choosing by community size, the choice is not this one.
• Breadth of the internal API. The Python projects got to several internal endpoints first; our own contributor notes tell us to read theirs before writing anything new.
• Node, and a browser. It needs Node 22+ and a Chromium it drives; the Python tools need Python and the same browser. Nobody here escapes the browser.

Where it wins

• A REST API with a published, valid OpenAPI 3.1 specification — all 56 operations named and typed, so a client generates in about ten seconds in any language.
• Citations as data. ask_question returns the quoted passage behind every marker, with its source name, as JSON — not markers to parse out of prose. How, and what the shape is.
• Retrieval that spends no generation allowance: search_passages returns the indexed passages for a query without generating anything.
• Multi-account rotation with automatic re-authentication, because the allowance is per account.
• Security handled in the open: three advisories found, fixed and published with their CVSS, and a build that fails on an unjustified exception.

The constraint all of them share

None of these is official. Google publishes no API for the consumer product, so every project here carries a logged-in Google session into endpoints Google never documented — which is why they all need a browser login, and why they all break together when something changes upstream. The three properties of that session, and the one official API that exists, are in does Gemini Notebook have an API?.

Pick by what you are building

If you are…	Pick
writing Python	notebooklm-py
wanting a command line at hand	gemini-notebook-mcp-cli
calling it from n8n, Zapier, Make, cron or curl	this project
wanting citations as structured data	this project
wanting a typed client in Rust or Ruby	this project
running several Google accounts	this project
already on the archived PleasePrompto server	migrate when convenient

Install is one command and one login.