NotebookLM-py
🚀 Description
teng-lin/notebooklm-py — an unofficial Python API + CLI + MCP server + agent skill for Google NotebookLM, giving programmatic access to features the web UI doesn’t expose. Uses undocumented Google endpoints — use at your own risk: not affiliated with Google, APIs can break without notice, rate limits apply. Best for prototypes, research, personal projects.
The core idea: NotebookLM is a grounded engine — Gemini reads your sources and answers with citations. Let it do the expensive server-side reasoning while your agent (Claude Code, Codex, …) orchestrates and handles the final mile. A zero-token synthesis + memory layer an agent drives in a loop, plus bulk structured export out.
🧩 Ways to use
| Method | Best for |
|---|---|
| CLI | shell scripts, quick tasks, CI/CD |
| Python API | app integration, async pipelines |
| MCP server | Claude Desktop/Code, Codex — local stdio or self-hosted remote connector (Cloudflare/Tailscale tunnel → claude.ai + ChatGPT, mobile included) |
| REST server | local automation over guarded HTTP routes |
| Agent skill | notebooklm skill install → ~/.claude/skills, ~/.agents/skills; or npx skills add teng-lin/notebooklm-py |
🧩 What it covers
- Sources — bulk import URLs, YouTube, PDFs/Word/EPUB/audio/video/images, Google Drive, pasted text; web/Drive research agents (fast/deep) with auto-import
- Chat — cited Q&A (
ask --json), personas, save answers/history as notes - Generation — Audio Overview (podcast), video, slide deck, infographic, quiz, flashcards, report, data table, mind map
- Beyond the web UI — batch downloads, quiz/flashcard export (JSON/MD/HTML), mind-map JSON extraction, data-table CSV, slide deck PPTX/PDF + per-slide revision, programmatic sharing
- Auth — interactive Playwright login, import cookies from a signed-in browser, or a durable master token that self-heals expired sessions unattended (servers, CI, remote MCP)
Reasoning for
The recipe worth stealing: knowledge distillation → permanent skill — run Deep Research or load a doc corpus, let Gemini condense it, bake the result into a SKILL.md the agent loads at startup (build once, reuse with zero runtime tokens). Related patterns: a “Master Brain” notebook as persistent cross-session memory queried from CLAUDE.md; a source-grounded troubleshooting oracle over a fast-moving tool’s docs; and Obsidian sync — run the CLI from the vault root so reports/mind-map JSON/transcripts land as files, with community skills resolving NotebookLM citations into [[wikilinks]].
Install: uv tool install "notebooklm-py[browser]" then notebooklm login (auto-downloads Chromium). On Windows plain pip install also works (not externally-managed).
Alternatives considered
- Open Notebook — self-hostable open-source NotebookLM clone (own inference, runs local via Ollama). NotebookLM-py is the opposite tack: drive the real hosted NotebookLM via undocumented APIs. Clone = own your stack; py = Gemini’s quality + features, at API-break risk.
- Standing up your own vector DB + embedding pipeline — the zero-infra alternative NotebookLM-py’s grounded-memory recipe replaces
- Manual clicking in the NotebookLM web app — no batch, no scriptable export, no mind-map JSON
🔗 Links
- Repo: https://github.com/teng-lin/notebooklm-py
- Docs: CLI Reference · Python API · MCP Guide
- Walkthroughs: Claude Code + NotebookLM (video)
🔗 Related notes
- Open Notebook — self-hosted NotebookLM clone (the inverse approach)
- Agent Skills — the skill packaging it ships as
- LLM Knowledge Bases — the distill-to-skill / grounded-memory pattern
- Obsidian — the knowledge-graph-sync recipe target
- Second Brain Design — PKM system this can feed
- Extending Claude Code — Tools for Its Blind Spots — compiled article featuring this tool
Template: tool