What it is
notebooklm-py lets you script Google NotebookLM, which Google renamed Gemini Notebook in July 2026; the README says the library drives the same service and works unchanged. You can create notebooks, add sources from URLs, YouTube, files and Drive, chat with them, run Deep Research, and generate audio overviews, videos, slide decks, quizzes, flashcards, reports and mind maps.
It comes as a Python library, a CLI, an MCP server (local over stdio or self-hosted as a remote connector) and an agent skill; notebooklm skill install puts the skill into Claude Code's skills folder. It is unofficial and works by calling the same undocumented internal APIs that the web app uses.
Our review
notebooklm-py is the furthest-reaching way to drive NotebookLM from code, and the least guaranteed. It has about 19,900 GitHub stars, released v0.8.5 on 12 October 2026, and only 19 open issues, which suggests a maintainer keeping pace with Google's changes rather than falling behind them.
Its reach goes beyond the web app: batch downloads of every artifact of a type, mind maps as JSON, quizzes and flashcards as JSON or Markdown, data tables as CSV and slide decks as editable PPTX. The pattern the README pushes is cost-shifting. Gemini reads 30 documents server-side and the agent spends its own tokens only on the final answer, which moves the bill from your API account to your Google subscription.
The weak point is authentication. The default path is a Playwright browser login (the first run downloads about 170 MB of Chromium) or cookies lifted from Chrome. For servers and CI there is a master-token login that mints fresh cookies unattended; the README calls it "a powerful full-account credential" and advises a dedicated Google account. Any agent or remote MCP connector holding that token can act as you across Google, not just in NotebookLM.
Capacity is set by Google, not the library. A free account gets 100 notebooks of 50 sources each; Google AI Pro raises that to 500 notebooks of 300 sources, and Google marks every limit as subject to change. The README warns heavy use may be throttled, and its own study-set recipe paces requests deliberately to avoid that. Against PleasePrompto/notebooklm-skill, which automates the browser and has not released since November 2025, this is the more capable and more actively maintained route, though both break when Google changes its front end.
Where it earns its place
A solo researcher using Claude Code on a Google AI Pro plan
Synthesise 30 papers without spending Claude tokens on reading them
The agent creates a notebook, adds the sources and asks NotebookLM questions; Gemini does the reading under the Pro plan's 300-source cap, and Claude only writes up the cited answers.
A platform engineer at a 30-person startup
Give coding agents a cited answer source for internal RFCs and architecture docs
A notebook of the docs is exposed through the MCP server, avoiding a vector database. It runs on one Google login, so a dedicated account and a short-lived cookie session are safer than the master token.
A training lead at a 200-person firm
Turn onboarding material into podcasts, quizzes and flashcards for each topic
A script builds one notebook per topic and bulk-exports quizzes as JSON and audio as MP3. Generation counts against the account's daily quotas, and requests need pacing to avoid throttling.
Before you commit
- Setup
- uv tool install "notebooklm-py[browser]" then notebooklm login, which downloads about 170 MB of Chromium for Google sign-in on first run; Python 3.10 or newer. Embedding the library without Playwright is a plain pip install.
- Integrations
- Python async API, CLI, an MCP server for Claude Desktop, Claude Code and Codex, a self-hostable remote connector for claude.ai and ChatGPT, and a skill installed with notebooklm skill install or npx skills add.
- Data and privacy
- MIT licence, no telemetry mentioned in the README. Sources and chats go to Google under your account; Google says personal-account NotebookLM content is not used to train its models unless you submit feedback, and Workspace accounts get stronger terms. Cookies or a master token are stored in a local profile.
- Cost at scale
- The library is free; capacity is your Google plan. Free accounts get 100 notebooks of 50 sources, Google AI Plus 200 of 100, Google AI Pro 500 of 300 and Ultra up to 600 sources per notebook.
- Limits
- Not affiliated with Google, and endpoints can change without notice; heavy use may be throttled. 19 open issues at the time of checking.
What's good
- Exposes exports the web app lacks: batch downloads, mind-map JSON, PPTX slides
- Library, CLI, MCP server and Claude Code skill in one package
- Frequent releases and a short issue list
What's not
- Undocumented Google APIs can change without notice
- Master-token login is a full Google account credential
- Throttling on heavy use; notebook and source caps set by your Google tier
Best for
Researchers who want Claude Code to build and query NotebookLM notebooks.
Alternatives
notebooklm-py: common questions
- Is notebooklm-py worth it?
- NotebookLM's full feature set in an agent's hands, held together by undocumented Google endpoints and a login you should not give it lightly. We rate it 3.8 out of 5.
- How much does notebooklm-py cost?
- Free and open source (MIT). NotebookLM limits depend on your Google account tier.
- What are the downsides of notebooklm-py?
- Undocumented Google APIs can change without notice; master-token login is a full Google account credential; throttling on heavy use; notebook and source caps set by your Google tier.
- Who is notebooklm-py best for?
- Researchers who want Claude Code to build and query NotebookLM notebooks.
- How much does notebooklm-py cost for a team or in production?
- The library is free; capacity is your Google plan. Free accounts get 100 notebooks of 50 sources, Google AI Plus 200 of 100, Google AI Pro 500 of 300 and Ultra up to 600 sources per notebook.
- Is notebooklm-py safe for company data?
- MIT licence, no telemetry mentioned in the README. Sources and chats go to Google under your account; Google says personal-account NotebookLM content is not used to train its models unless you submit feedback, and Workspace accounts get stronger terms. Cookies or a master token are stored in a local profile.
- What are the best notebooklm-py alternatives?
- The closest alternatives we have reviewed are claude-obsidian and claude-mem.
Answers checked Oct 23, 2026.