notebooklm-py AI工具

Or use Microsoft Edge (for orgs that require Edge for SSO)

What You Can Build

🤖 AI Agent Tools - Integrate NotebookLM into Claude Code, Codex, and other LLM agents. Ships with a root NotebookLM skill for GitHub and npx skills add discovery, local notebooklm skill install support for Claude Code and .agents skill directories, and repo-level Codex guidance in AGENTS.md.

📚 Research Automation - Bulk-import sources (URLs, PDFs, YouTube, Google Drive), run web/Drive research queries with auto-import, and extract insights programmatically. Build repeatable research pipelines.

🎙️ Content Generation - Generate Audio Overviews (podcasts), videos, slide decks, quizzes, flashcards, infographics, data tables, mind maps, and study guides. Full control over formats, styles, and output.

📥 Downloads & Export - Download all generated artifacts locally (MP3, MP4, PDF, PNG, CSV, JSON, Markdown). Export to Google Docs/Sheets. Features the web UI doesn't offer: batch downloads, quiz/flashcard export in multiple formats, mind map JSON extraction.

Installation

The full install guide — six personas (agent, end-user, library, headless, contributor, power-user), optional extras matrix, platform notes — lives in docs/installation.md.

Quickest start (CLI users and AI agents) — install the CLI with uv tool (recommended) or pipx:

uv tool install "notebooklm-py[browser]"   # or: pipx install "notebooklm-py[browser]"
notebooklm login                           # first run auto-downloads Chromium (~170 MB), then Google sign-in
notebooklm auth check --test --json        # verify: expect "status": "ok"

Why uv tool / pipx? They install the CLI into its own isolated environment and put notebooklm on your PATH — no dependency clashes with other tools, a one-line upgrade (uv tool upgrade notebooklm-py) or uninstall, and, crucially, they work on modern macOS (Homebrew Python) and Debian/Ubuntu where a system-wide pip install is blocked with error: externally-managed-environment (PEP 668). No uv yet? curl -LsSf https://astral.sh/uv/install.sh | sh (or brew install uv / winget install astral-sh.uv).

Prefer plain pip? It works the same inside a virtualenv (and directly on Windows, where Python isn't externally-managed):

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install "notebooklm-py[browser]"

As a library (embedded in your app — no Playwright, no Chromium):

uv add notebooklm-py                    # or, inside a virtualenv: pip install notebooklm-py

If playwright install chromium fails on Linux with TypeError: onExit is not a function, see the Linux workaround. Contributors: see CONTRIBUTING.md.

Agent Setup

Option 1 — CLI install:

notebooklm skill install

Installs the skill into ~/.claude/skills/notebooklm and ~/.agents/skills/notebooklm.

Option 2 — npx install (via the open skills ecosystem):

npx skills add teng-lin/notebooklm-py

Fetches the canonical SKILL.md directly from GitHub.

Use Cases & Recipes

NotebookLM is a grounded engine: Gemini does the heavy reading and answers from your sources with citations. The winning pattern is to let it do the expensive analysis while your agent (Claude Code, Codex, …) orchestrates and handles the final mile. Recipes people build on top of this library:

• 🪙 Zero-token research offload — Throw 30 documents into a notebook, let Gemini do the heavy analysis, and have your agent spend tokens only on the final polish. The agent just orchestrates (create → source add → ask); the reasoning happens server-side.
• 🧠 Web research → expert agent — Run Deep Research (source add-research) to scan the web into a sourced report, then distill that report into a reusable Claude skill — a packaged domain expert without hand-curating sources.
• 💾 Persistent cross-session memory — Keep a "Master Brain" notebook; a wrap-up step appends each session's decisions and fixes as notes (note create / ask --save-as-note), and a line in your CLAUDE.md queries it (ask) at the start of the next session. Storage and recall live on Google's infrastructure.
• 🕸️ Obsidian / knowledge-graph sync — Run the CLI from your vault root so downloaded artifacts (reports, mind-map JSON, transcripts) land as files in your knowledge graph; community skills built on this library even resolve NotebookLM's citation markers into Obsidian [[wikilinks]]. Pair with a podcast overview for an audio digest of your notes.
• 🔁 Multi-format content repurposing — One source set, every format: generate audio (podcast), generate video, generate slide-deck, plus a generate report blog draft, generate quiz, and generate flashcards — fan a single notebook out across channels.
• 📞 Grounded knowledge base (RAG) — Load product docs, FAQs, RFCs, and past tickets, then ask --json for source-grounded, cited answers for support, on-call, or internal Q&A.
• 🧩 Grounded memory for coding agents — Expose a notebook of your internal docs/RFCs/architecture over the MCP server (or plain ask) so an agent answers from your code with citations rather than plausible-sounding guesses — a zero-infra alternative to standing up your own vector DB and embedding pipeline.
• 🚨 Incident runbook generator — On an alert, spin up a notebook of the relevant docs, ask targeted diagnostic questions, and emit a briefing-doc report (generate report --format briefing-doc) as an automated runbook.
• 📚 Curriculum / study-set builder — Scrape a syllabus or developer roadmap, create one notebook per topic (with deliberate pacing to dodge rate limits), and bulk-generate podcasts, quizzes, and flashcards for each.
• 📰 Scheduled audio briefings — Pair auth refresh --quiet (cron/launchd/systemd) with generate audio to publish a fresh personalized briefing to a podcast feed on a schedule.

These combine ordinary library primitives — see the CLI Reference and Python API. The agent-side glue (skills, scheduling, vault layout) lives in your own setup, not this package.

Seen in the wild: "Claude Code + NotebookLM = CHEAT CODE" · "…+ Obsidian = GOD MODE" · a browser-free YouTube→notebook→cited-answers pipeline driven entirely from the terminal · a four-workflow guide to offloading heavy document analysis onto NotebookLM so Claude Code stops burning tokens (zero-token research, web-research agents, cross-session memory, an Obsidian "second brain") · turning a notebook into the source-grounded "project brain" a coding agent consults before it writes code.

Quick Start

<p align="center"> <a href="https://asciinema.org/a/767284" target="_blank"><img src="https://asciinema.org/a/767284.svg" width="600" /></a> <br> <em>16-minute session compressed to 30 seconds</em> </p>

Python API

import asyncio
from notebooklm import NotebookLMClient, MindMapKind

async def main():
    async with NotebookLMClient.from_storage() as client:
        # Create notebook and add sources
        nb = await client.notebooks.create("Research")
        await client.sources.add_url(nb.id, "https://example.com", wait=True)

        # Chat with your sources
        result = await client.chat.ask(nb.id, "Summarize this")
        print(result.answer)

        # Generate content (podcast, video, quiz, etc.)
        status = await client.artifacts.generate_audio(nb.id, instructions="make it fun")
        await client.artifacts.wait_for_completion(nb.id, status.task_id)
        await client.artifacts.download_audio(nb.id, "podcast.mp3")

        # Generate quiz and download as JSON
        status = await client.artifacts.generate_quiz(nb.id)
        await client.artifacts.wait_for_completion(nb.id, status.task_id)
        await client.artifacts.download_quiz(nb.id, "quiz.json", output_format="json")

        # Generate a mind map via the unified client.mind_maps API (issue #1256) —
        # two kinds: the newer MindMapKind.INTERACTIVE studio map (shown; polled to
        # completion by default) or MindMapKind.NOTE_BACKED JSON. Both export via:
        await client.mind_maps.generate(nb.id, kind=MindMapKind.INTERACTIVE)
        await client.artifacts.download_mind_map(nb.id, "mindmap.json")

asyncio.run(main())