notebooklm-mcp

An unofficial NotebookLM MCP server and CLI — inspired by jacob-bd/notebooklm-mcp-cli.

⚠️ NotebookLM has no public API. This project talks to NotebookLM's internal batchexecute / GenerateFreeFormStreamed RPCs through a Playwright persistent-context session. It is not affiliated with Google.

What works today

Tool	Status	Notes
notebook_list	✅	Returns "mine" + "shared" notebooks. Pagination for many owned notebooks not yet discovered (shows most-recent owned).
notebook_create	✅	Untitled empty notebook.
notebook_delete	✅	Permanent, no undo.
notebook_query	✅	Streams the answer with inline [1][2] citations.
source_add	⚠️	kind="text" and kind="url" work. drive / file stubs.
research_start	✅	NotebookLM Discover → returns candidate URLs (does not auto-import).
research_and_ask	✅	One-shot: create → research → import → ask N questions → (optional) delete.
studio_create	⚠️	artifact="audio" (podcast) works. video / slides / mindmap / etc. stubs.
download_artifact	✅	Saves completed audio as .m4a.
refresh_auth	✅	Checks whether the stored Google session is still live.

Stubs (raise NotImplementedError): notebook_share_public, notebook_share_invite, source_sync_drive, source_get_content, studio_revise, cross_notebook_query, batch, pipeline, tag.

End-to-end workflow this supports

Manual orchestration:

notebook_create()              → empty notebook
research_start(nid, topic)     → ~10 candidate URLs
source_add(nid, kind="url")×N  → import the ones you want
notebook_query(nid, question)  → grounded answer with citations
studio_create(nid, "audio")    → start podcast generation (2-5 min)
download_artifact(nid)         → .m4a file on disk

Or, in a single call:

research_and_ask(
    topic="history of the Cold War space race",
    questions=["Who reached space first?", "Key consequences?"],
    max_sources=3,
    keep_notebook=False,  # ephemeral — delete after
)
# → { sources_imported: [...], qa: [{question, answer, ...}, ...] }

Install

Quick start (from PyPI)

uv tool install notebooklm-mcp-lisa
uvx --from notebooklm-mcp-lisa playwright install chromium
nlm login                       # one-time Google sign-in (opens a browser)
nlm setup add claude-code       # or: claude-desktop

The setup add command edits the target's config file so the MCP server is registered — no JSON editing by hand. Restart the client to pick it up.

Verify registration any time:

nlm setup list

Dev setup (from source)

git clone https://github.com/gracelee087/notebooklm-mcp-lisa.git
cd notebooklm-mcp
uv sync
uv run playwright install chromium
uv run nlm login

First-time Google login

NotebookLM requires a signed-in Google session. Run once, headed:

uv run nlm login

A Chromium window opens. Complete Google OAuth; the session is saved to the OS user-data dir (NLM_PROFILE_DIR to override). Subsequent runs are headless.

Verify:

uv run nlm status
# {'logged_in': True}

Run the MCP server

uv run notebooklm-mcp
# or
uv run nlm serve

Claude Code (project-scoped, auto-picked-up)

.mcp.json in this repo is already configured. When you open the project in Claude Code it prompts you to trust and loads notebooklm automatically.

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "notebooklm": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/this/repo", "run", "notebooklm-mcp"]
    }
  }
}

Repo layout

src/notebooklm_mcp/
  server.py          # FastMCP entry
  cli.py             # `nlm` CLI (login/status/serve)
  browser.py         # Playwright persistent context
  config.py          # paths, env
  tools/
    notebook.py source.py studio.py research.py workflow.py auth.py

Status

• [x] Package scaffold, MCP server, CLI
• [x] Playwright session with persistent Google login
• [x] All tools registered (stubs)
• [ ] notebook_list DOM scrape
• [ ] notebook_query chat flow
• [ ] source_add (url/text/drive/file)
• [ ] studio_create audio (podcast) + download
• [ ] remaining tools

License

MIT