The Problem

When you tell Claude Code or Cursor to "search through my local documentation", here's what happens:

• Massive token consumption: Searching through documentation means reading multiple files repeatedly
• Inaccurate retrieval: Searches for keywords, misses context and connections between docs
• Hallucinations: When it can't find something, it invents plausible-sounding APIs
• Expensive & slow: Each question requires re-reading multiple files

The Solution

Let your local agents chat directly with NotebookLM — Google's zero-hallucination knowledge base powered by Gemini 2.5 that provides intelligent, synthesized answers from your docs.

Your Task → Local Agent asks NotebookLM → Gemini synthesizes answer → Agent writes correct code

The real advantage: No more manual copy-paste between NotebookLM and your editor. Your agent asks NotebookLM directly and gets answers straight back in the CLI. It builds deep understanding through automatic follow-ups — Claude asks multiple questions in sequence, each building on the last, getting specific implementation details, edge cases, and best practices. You can save NotebookLM links to your local library with tags and descriptions, and Claude automatically selects the relevant notebook based on your current task.

Why NotebookLM, Not Local RAG?

What Makes NotebookLM Superior?

1. Pre-processed by Gemini: Upload docs once, get instant expert knowledge
2. Natural language Q&A: Not just retrieval — actual understanding and synthesis
3. Multi-source correlation: Connects information across 50+ documents
4. Citation-backed: Every answer includes source references
5. No infrastructure: No vector DBs, embeddings, or chunking strategies needed

Installation

Claude Code

claude mcp add notebooklm npx notebooklm-mcp@latest

Codex

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

gemini mcp add notebooklm npx notebooklm-mcp@latest

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

amp mcp add notebooklm -- npx notebooklm-mcp@latest

code --add-mcp '{"name":"notebooklm","command":"npx","args":["notebooklm-mcp@latest"]}'

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

Alternative: Claude Code Skill

Prefer Claude Code Skills over MCP? This server is now also available as a native Claude Code Skill with a simpler setup:

NotebookLM Claude Code Skill - Clone to ~/.claude/skills and start using immediately

Key differences:

• MCP Server (this repo): Persistent sessions, works with Claude Code, Codex, Cursor, and other MCP clients
• Claude Code Skill: Simpler setup, Python-based, stateless queries, works only with local Claude Code

Both use the same browser automation technology and provide zero-hallucination answers from your NotebookLM notebooks.

Quick Start

1. Install the MCP server (see Installation above)

2. Authenticate (one-time)

Say in your chat (Claude/Codex):

"Log me in to NotebookLM"

A Chrome window opens → log in with Google

3. Create your knowledge base

Go to notebooklm.google.com → Create notebook → Upload your docs:

• 📄 PDFs, Google Docs, markdown files
• 🔗 Websites, GitHub repos
• 🎥 YouTube videos
• 📚 Multiple sources per notebook

Share: ⚙️ Share → Anyone with link → Copy

4. Let Claude use it

"I'm building with [library]. Here's my NotebookLM: [link]"

That's it. Claude now asks NotebookLM whatever it needs, building expertise before writing code.

Real-World Example

Building an n8n Workflow Without Hallucinations

Challenge: n8n's API is new — Claude hallucinates node names and functions.

Solution:

1. Downloaded complete n8n documentation → merged into manageable chunks
2. Uploaded to NotebookLM
3. Told Claude: "Build me a Gmail spam filter workflow. Use this NotebookLM: [link]"

Watch the AI-to-AI conversation:

Claude → "How does Gmail integration work in n8n?"
NotebookLM → "Use Gmail Trigger with polling, or Gmail node with Get Many..."

Claude → "How to decode base64 email body?"
NotebookLM → "Body is base64url encoded in payload.parts, use Function node..."

Claude → "How to parse OpenAI response as JSON?"
NotebookLM → "Set responseFormat to json, use {{ $json.spam }} in IF node..."

Claude → "What about error handling if the API fails?"
NotebookLM → "Use Error Trigger node with Continue On Fail enabled..."

Claude → ✅ "Here's your complete workflow JSON..."

Result: Perfect workflow on first try. No debugging hallucinated APIs.

Core Features

Zero Hallucinations

NotebookLM refuses to answer if information isn't in your docs. No invented APIs.

Autonomous Research

Claude asks follow-up questions automatically, building complete understanding before coding.

Smart Library Management

Save NotebookLM links with tags and descriptions. Claude auto-selects the right notebook for your task.

"Add [link] to library tagged 'frontend, react, components'"

Deep, Iterative Research

• Claude automatically asks follow-up questions to build complete understanding
• Each answer triggers deeper questions until Claude has all the details
• Example: For n8n workflow, Claude asked multiple sequential questions about Gmail integration, error handling, and data transformation

Cross-Tool Sharing

Set up once, use everywhere. Claude Code, Codex, Cursor — all share the same library.

Deep Cleanup Tool

Fresh start anytime. Scans entire system for NotebookLM data with categorized preview.

Architecture

graph LR
    A[Your Task] --> B[Claude/Codex]
    B --> C[MCP Server]
    C --> D[Chrome Automation]
    D --> E[NotebookLM]
    E --> F[Gemini 2.5]
    F --> G[Your Docs]
    G --> F
    F --> E
    E --> D
    D --> C
    C --> B
    B --> H[Accurate Code]

Common Commands

Comparison to Alternatives

vs. Downloading docs locally

• You: Download docs → Claude: "search through these files"
• Problem: Claude reads thousands of files → massive token usage, often misses connections
• NotebookLM: Pre-indexed by Gemini, semantic understanding across all docs

vs. Web search

• You: "Research X online"
• Problem: Outdated info, hallucinated examples, unreliable sources
• NotebookLM: Only your trusted docs, always current, with citations

vs. Local RAG setup

• You: Set up embeddings, vector DB, chunking strategy, retrieval pipeline
• Problem: Hours of setup, tuning retrieval, still gets "creative" with gaps
• NotebookLM: Upload docs → done. Google handles everything.

FAQ

Is it really zero hallucinations? Yes. NotebookLM is specifically designed to only answer from uploaded sources. If it doesn't know, it says so.

What about rate limits? Free tier has daily query limits per Google account. Quick account switching supported for continued research.

How secure is this? Chrome runs locally. Your credentials never leave your machine. Use a dedicated Google account if concerned.

Can I see what's happening? Yes! Say "Show me the browser" to watch the live NotebookLM conversation.

What makes this better than Claude's built-in knowledge? Your docs are always current. No training cutoff. No hallucinations. Perfect for new libraries, internal APIs, or fast-moving projects.

Advanced Usage

• 📖 Usage Guide — Patterns, workflows, tips
• 🛠️ Tool Reference — Complete MCP API
• 🔧 Configuration — Environment variables
• 🐛 Troubleshooting — Common issues

The Bottom Line

Without NotebookLM MCP: Write code → Find it's wrong → Debug hallucinated APIs → Repeat

With NotebookLM MCP: Claude researches first → Writes correct code → Ship faster

Stop debugging hallucinations. Start shipping accurate code.

# Get started in 30 seconds
claude mcp add notebooklm npx notebooklm-mcp@latest

Disclaimer

This tool automates browser interactions with NotebookLM to make your workflow more efficient. However, a few friendly reminders:

About browser automation: While I've built in humanization features (realistic typing speeds, natural delays, mouse movements) to make the automation behave more naturally, I can't guarantee Google won't detect or flag automated usage. I recommend using a dedicated Google account for automation rather than your primary account—think of it like web scraping: probably fine, but better safe than sorry!

About CLI tools and AI agents: CLI tools like Claude Code, Codex, and similar AI-powered assistants are incredibly powerful, but they can make mistakes. Please use them with care and awareness:

• Always review changes before committing or deploying
• Test in safe environments first
• Keep backups of important work
• Remember: AI agents are assistants, not infallible oracles

I built this tool for myself because I was tired of the copy-paste dance between NotebookLM and my editor. I'm sharing it in the hope it helps others too, but I can't take responsibility for any issues, data loss, or account problems that might occur. Use at your own discretion and judgment.

That said, if you run into problems or have questions, feel free to open an issue on GitHub. I'm happy to help troubleshoot!

Contributing

Found a bug? Have a feature idea? Open an issue or submit a PR!

License

MIT — Use freely in your projects.