notebooklm-mcp

notebooklm-mcp

A secure MCP (Model Context Protocol) server for Google NotebookLM. Enables AI assistants to interact with NotebookLM notebooks — manage notebooks, query sources, create audio/video/reports, run research, and more.

v2 uses direct HTTP API calls (no browser automation at runtime), making it significantly faster than v1 while maintaining all security hardening.

Security

Issue in existing MCPs	Our fix
Plaintext cookie storage	AES-256-GCM encrypted, key in macOS Keychain
No file permissions	0o700 dirs, 0o600 files, verified on read
--no-sandbox Chrome	Sandbox enabled; --no-sandbox in blocklist
Remote debug port exposed	--remote-debugging-port in blocklist
Stored Google password	Manual login only — no credential storage
No URL validation	All URLs validated against notebooklm.google.com
Prompt injection via DOM	Response sanitizer strips zero-width chars, flags injection patterns
All cookies stored	Only .google.com + notebooklm.google.com persisted
exec() for shell commands	execFile() to prevent shell injection

Tools (35)

Authentication (3)

Tool	Description
setup_auth	Interactive login (headful browser, one-time setup)
check_auth	Validate auth status
clear_auth	Clear all stored auth data

Notebook Management (7)

Tool	Description
notebook_list	List all notebooks in your account
notebook_create	Create a new notebook
notebook_get	Get notebook details including sources
notebook_describe	Get AI-generated summary and suggested topics
notebook_rename	Rename a notebook
notebook_delete	Delete a notebook (requires confirm=true)
chat_configure	Configure AI chat settings (goal, response length, custom prompt)

Source Management (8)

Tool	Description
source_add_url	Add a URL or YouTube link as source
source_add_text	Add text content as source
source_add_drive	Add a Google Drive document as source
source_describe	Get AI-generated source summary
source_get_content	Get original text content of a source
source_list_drive	List Drive sources with sync status
source_sync_drive	Sync a Drive source to latest version
source_delete	Delete a source (requires confirm=true)

AI Query (1)

Tool	Description
notebook_query	Ask questions about notebook sources (supports follow-ups)

Research (3)

Tool	Description
research_start	Start web research (fast or deep mode)
research_status	Check research task progress
research_import	Import research results as notebook source

Studio Content (10)

Tool	Description
audio_create	Generate audio podcast (format, length, language, focus)
video_create	Generate video overview (visual style, language, focus)
report_create	Generate report (briefing doc, study guide, FAQ, timeline, blog)
flashcards_create	Generate flashcards (difficulty)
quiz_create	Generate quiz (question count, difficulty)
infographic_create	Generate infographic (orientation, detail level)
slide_deck_create	Generate slide deck (format, length)
data_table_create	Generate data table
studio_status	Check generation status of any studio artifact
studio_delete	Delete a studio artifact (requires confirm=true)

Mind Map (3)

Tool	Description
mind_map_create	Generate a mind map from sources (synchronous, returns immediately)
mind_map_list	List all mind maps in a notebook
mind_map_delete	Delete a mind map (requires confirm=true)

Requirements

• Node.js 18+
• macOS (uses Keychain for key storage)
• Google account with NotebookLM access

Installation

git clone https://github.com/ers123/notebooklm-mcp.git
cd notebooklm-mcp
npm install
npm run build
npx patchright install chromium   # needed for one-time setup_auth only

Configuration

Claude Code

Add to ~/.claude/mcp_config.json:

{
  "mcpServers": {
    "notebooklm": {
      "command": "node",
      "args": ["/path/to/notebooklm-mcp/dist/index.js"]
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "notebooklm": {
      "command": "node",
      "args": ["/path/to/notebooklm-mcp/dist/index.js"]
    }
  }
}

Any MCP Client (stdio)

node /path/to/notebooklm-mcp/dist/index.js

Usage

First-time setup

1. Connect the MCP server to your AI assistant
2. Run setup_auth — a browser window opens
3. Sign in to your Google account manually
4. The browser closes automatically after successful login
5. Cookies are encrypted and stored securely
6. All subsequent API calls use HTTP directly (no browser needed)

Example workflows

# List your notebooks
Use notebook_list

# Create a notebook and add sources
Use notebook_create with title "AI Research"
Use source_add_url with notebookId "..." and url "https://example.com/paper.pdf"
Use source_add_text with notebookId "..." and title "Notes" and content "..."

# Query your notebook
Use notebook_query with notebookId "..." and question "What are the key findings?"
Use notebook_query with notebookId "..." and question "Tell me more" and followUp true

# Generate content
Use audio_create with notebookId "..." and format "conversation" and length "medium"
Use report_create with notebookId "..." and reportFormat "briefing_doc"
Use studio_status with notebookId "..." and artifactId "..." to check progress

# Run web research
Use research_start with notebookId "..." and query "latest AI developments" and mode "deep"
Use research_status with notebookId "..." and taskId "..." to check progress
Use research_import with notebookId "..." and taskId "..." to add results as source

Architecture

src/
├── index.ts                    # Entry point + stdio transport
├── server.ts                   # 35 MCP tools registered
├── config.ts                   # Constants, API endpoints, timeouts
├── types.ts                    # Shared interfaces
├── errors.ts                   # Error class hierarchy
├── api/                        # HTTP API layer (v2 core)
│   ├── rpc-client.ts           # batchexecute RPC client
│   ├── query-client.ts         # Streaming query client
│   ├── response-parser.ts      # Anti-XSSI removal + chunk parsing
│   └── constants.ts            # RPC IDs + code mappings
├── security/                   # Keychain, AES-256-GCM, file perms, URL validation, sanitizer
├── auth/                       # Manual login flow, encrypted cookie store
├── browser/                    # Patchright launcher (setup_auth only)
└── tools/                      # 35 MCP tool handlers
    ├── notebook/               # list, create, get, describe, rename, delete, chat-configure
    ├── source/                 # add-url, add-text, add-drive, describe, get-content, list-drive, sync-drive, delete
    ├── query/                  # notebook-query (AI Q&A with follow-ups)
    ├── research/               # start, status, import
    ├── studio/                 # audio, video, report, flashcards, quiz, infographic, slide-deck, data-table, mind-map, status, delete
    └── auth-tools/             # setup, check, clear auth

Development

npm run dev          # Run with tsx (hot reload)
npm run build        # Compile TypeScript
npm test             # Run unit tests
npm run test:watch   # Run tests in watch mode

Dependencies

Only 2 runtime dependencies:

• @modelcontextprotocol/sdk — MCP server framework
• zod — Input validation

Optional:

• patchright — Browser automation (only needed for setup_auth)

No native modules. Encryption via Node.js built-in crypto. Keychain via child_process.execFile.

Acknowledgments

The NotebookLM internal API structure (RPC IDs, parameter formats, response schemas) was informed by community reverse-engineering efforts, including wonseokjung/notebooklm-mcp. All implementation code is original.

License

MIT