NotebookLM MCP Server

I fell in love with NotebookLM because it’s one of the best “RAG-style” research companions I’ve used. I used to upload docs, ask NotebookLM questions, then copy the answers back into Codex/Claude Code by hand. It worked, but it was clunky. So I built notebooklm-mcp so my code agents can talk to NotebookLM directly—no copy/paste, just grounded answers right inside the IDE chat.

Why you might want this

• Grounded answers on demand – the server drives a real Chrome session, so Gemini always cites from your NotebookLM notebooks.

• Agent-first flow – speak naturally to Claude/Codex; it calls the MCP tools and every answer ends with a reminder to keep asking until the task is complete.

• No hallucinations – NotebookLM only answers with what’s in your notebooks. Feed it good documentation and you’ll get accurate output.

• Library management – add, update, remove, or search notebooks without leaving chat.

• Persistent auth – Chrome profile keeps cookies/sessions; switch Google accounts whenever you like.

Free NotebookLM accounts allow 50 chat turns per day. Hit the cap? Re-run auth setup with another Google account and keep going.

💡 Cross-Client Sharing: All MCP clients (Claude Code, Codex, etc.) share the same authentication and notebook library. Authenticate once, add notebooks once—switch between clients without re-login or re-adding notebooks.

Preparation checklist

1. Grab a Google account (throwaway is fine).

2. Visit https://notebooklm.google/ and click Try NotebookLM.

3. Upload your docs via the New button. NotebookLM lets you store 100 notebooks, each with up to 50 sources and roughly 500,000 words—plenty for merged documentation sets.

4. Share the notebook: Click Share (top right) → change Notebook access from "Restricted" to "Anyone with the link" → optionally restrict Viewer permissions to "Chat only" → click Copy link. Give this link to Claude/Codex so it can add the notebook to your library.

Quick tour

Every ask_question reply ends with:

EXTREMELY IMPORTANT: Is that ALL you need to know? … so the agent remembers to plan follow-up questions before it answers you.

Conversation-led quick start

Conversation tips (inspired by the n8n workflow test)

• State the goal and share the notebook upfront. Claude instantly knew to add the n8n documentation notebook before researching.

• Confirm library changes. The agent asks before adding/updating/removing notebooks—say “yes/no” so your library stays tidy.

• Let the agent interrogate NotebookLM. Each answer ends with the reminder to keep asking questions; encourage it to follow through before replying to you.

• Keep sessions scoped. Start a new session for a new project so context doesn’t leak.

• Review the output. Spot-check JSON, node types, connections—if something’s off, the agent can recheck the doc and fix it.

• Feed NotebookLM quality sources. The answers are only as good as the docs you upload—no hallucinations, just your knowledge base.

1. Install the MCP server

   Claude Code:

   claude mcp add notebooklm npx notebooklm-mcp@latest

   Codex:

   codex mcp add notebooklm -- npx notebooklm-mcp@latest

   Other MCP clients (manual config):

   { "mcpServers": { "notebooklm": { "command": "npx", "args": ["notebooklm-mcp@latest"] } } }

2. Authenticate through chat

   • Ask your agent to “open NotebookLM auth setup”.

   • Chrome opens; sign in with the Google account you want to use.

   • When NotebookLM loads, close the window. Optionally ask it to run get_health to confirm.

3. Add notebooks as you go

   • Paste a NotebookLM link: “Add https://notebooklm.google.com/notebook/abc123 to our library.”

   • The agent confirms metadata before writing to the library.

4. Research before answering

   • “Before you answer, research the n8n Gmail spam workflow in NotebookLM.”

   • The agent keeps the same session_id, asks successive questions, reads the reminder, and only then replies.

5. Reauthenticate whenever needed

   • “Reauthenticate with a different Google account.”

   • The server clears the profile, opens Chrome, and you sign in again.

How it works under the hood

• Real Chrome instance via Playwright (Patchright) keeps cookies, fingerprint, and session storage persistent.

• SessionManager enforces timeouts, max sessions, and auto cleanup.

• Library metadata (library.json) lives under ${CONFIG.dataDir}.

• Browser profile lives under ${CONFIG.chromeProfileDir}; delete it to reset auth.

• All tools are exposed through MCP—no built-in prompts or slash commands.

• NotebookLM only answers from the notebooks you supply. If the info isn’t there, Gemini won’t invent it.

Data Storage & Cross-Platform Support

Shared State Architecture: All MCP clients (Claude Code, Codex, etc.) share the same authentication, notebooks, and browser state. Authenticate once, add notebooks once, and they're available everywhere.

Storage Locations

All data is stored in platform-specific directories using env-paths:

What's Stored

• Chrome Profile (chrome_profile/) – Browser cookies, session storage, fingerprint (persistent across restarts)

• Notebook Library (library.json) – All notebooks you've added with metadata

• Browser State (browser_state/) – Auth snapshots (legacy/fallback)

• Isolated Instances (chrome_profile_instances/) – Temporary profiles for concurrent MCP server instances

Platform Support

• Linux: ✅ Fully tested and production-ready

• macOS: Should work (dependencies support it, but not extensively tested)

• Windows: Untested (technically compatible, but no guarantees)

FAQ

Can I trust the auth flow? Yes. Chrome runs locally, and the server never handles your Google credentials beyond Playwright's APIs. Use a dedicated account if you prefer.

What happens when NotebookLM hits the 50-query limit? You'll see a rate-limit error. Ask the agent to reopen auth setup with another account or wait until the next day. The free tier allows 50 chat turns/day—swap accounts via auth setup when you hit the cap.

Do I need to manage session IDs manually? Claude/Codex usually keep it for you. If you call tools manually, capture the first session_id and reuse it for follow-ups.

Can I show the browser while the agent runs? Yes—set show_browser: true on ask_question, or ask the agent to toggle it for debugging.

Does the server still ship prompts or slash commands? No. Everything is driven by natural-language instructions.

How do I keep the agent on track? Share the notebook link, grant permission for library edits, and remind it to plan with NotebookLM before replying—just like in the n8n workflow example above.

What if Chrome crashes? The next ask_question recreates the browser context automatically.

What if authentication fails? Ask the agent to repair auth—it will run get_health, reopen Chrome, and verify again.

Should I back up my data? Yes—back up library.json (see Data Storage section above for location) if you switch machines or want to preserve your notebook library.

Configuration highlights

See docs/configuration.md for the full list (profile strategy, typing speeds, cleanup settings, etc.).

Documentation

• docs/quickstart.md – conversational setup.

• docs/tools.md – tool/resource reference.

• docs/troubleshooting.md – common issues & fixes.

• NOTEBOOKLM_USAGE.md – chat scenarios and phrases.

License

MIT