Skip to content

GEMINI.md

Latest commit

 

History

History
140 lines (107 loc) · 5.97 KB

GEMINI.md

File metadata and controls

140 lines (107 loc) · 5.97 KB

GEMINI.md

Project Overview

NotebookLM MCP Server & CLI

This project implements a Model Context Protocol (MCP) server and a full-featured Command Line Interface (CLI) that provides programmatic access to NotebookLM. It allows AI agents, developers, and power users to interact with NotebookLM notebooks, sources, query capabilities, and download generated artifacts (Audio, Video, PDF, etc.).

Tested with personal/free tier accounts. May work with Google Workspace accounts but has not been tested. This project relies on internal APIs (batchexecute RPCs).

Environment & Setup

The project uses uv for dependency management and tool installation.

Prerequisites

  • Python 3.11+
  • uv (Universal Python Package Manager)
  • Google Chrome (for automated authentication)

Installation

From PyPI (Recommended):

uv tool install notebooklm-mcp-cli
# or: pip install notebooklm-mcp-cli

From Source (Development):

git clone https://github.com/YOUR_USERNAME/notebooklm-mcp.git
cd notebooklm-mcp
uv tool install .

Authentication

Preferred: Run the automated authentication CLI:

nlm login

This launches Chrome, you log in, and cookies are extracted automatically. Your login is saved to a Chrome profile for future use.

Auto-refresh (v0.1.9+): The server now automatically handles token expiration:

  1. Refreshes CSRF tokens on expiry (immediate)
  2. Reloads cookies from disk if updated externally
  3. Runs headless Chrome auth if profile has saved login

If headless auth fails (Google login fully expired), you'll see a message to run nlm login again.

Explicit refresh (MCP tool):

refresh_auth()  # Reload tokens from disk or run headless auth

Fallback: Manual extraction (if CLI fails) If the automated tool doesn't work, extract cookies via Chrome DevTools:

  1. Open Chrome DevTools on notebooklm.google.com
  2. Go to Network tab, find a batchexecute request
  3. Copy the Cookie header and call save_auth_tokens(cookies=...)

Environment variable (advanced):

export NOTEBOOKLM_COOKIES="SID=xxx; HSID=xxx; SSID=xxx; ..."

Cookies last for weeks. The server auto-refreshes as long as Chrome profile login is valid.

Development Workflow

Building and Running

Reinstalling after changes: Because uv tool install installs into an isolated environment, you must reinstall to see changes during development.

uv cache clean
uv tool install --force .

Running the Server:

# Standard mode (stdio)
notebooklm-mcp

# Debug mode (verbose logging)
notebooklm-mcp --debug

# HTTP Server mode
notebooklm-mcp --transport http --port 8000

Testing

Run the test suite using pytest via uv:

# Run all tests
uv run pytest

# Run a specific test file
uv run pytest tests/test_api_client.py

Project Structure

  • src/notebooklm_tools/
    • services/: Shared service layer (v0.3.0+) — Business logic, validation, error handling
      • errors.py: Custom error hierarchy (ServiceError, ValidationError, etc.)
      • chat.py, downloads.py, exports.py, notebooks.py, notes.py: Domain services
      • research.py, sharing.py, sources.py, studio.py: More domain services
      • batch.py, cross.py, pipeline.py, smart_select.py: Batch, cross-notebook, pipeline, and tagging services
    • cli/: CLI commands and formatting (thin wrapper delegating to services/)
    • mcp/: MCP Server implementation (thin wrapper delegating to services/)
      • tools/: Modular tool definitions (one file per domain)
      • server.py: Slim server facade (imports tools from modules)
    • core/client.py: Low-level internal API calls (no business logic).
    • core/constants.py: Single source of truth for all API code-name mappings.
    • core/auth.py: Handles token validation, storage, and loading.
    • utils/cdp.py: Chrome DevTools Protocol for cookie extraction and headless auth.
    • utils/: Configuration and browser utilities
  • tests/services/: Unit tests for all service modules (576+ tests)
  • CLAUDE.md: Contains detailed documentation on the internal RPC IDs and protocol specifics. Refer to this file for API deep dives.
  • pyproject.toml: Project configuration and dependencies.

Key Conventions

  • Internal APIs: This project relies on undocumented APIs. Changes to Google's internal API will break functionality.
  • RPC Protocol: The API uses Google's batchexecute protocol. Responses often contain "anti-XSSI" prefixes ()]}') that must be stripped.
  • Layering (v0.3.0+): cli/ and mcp/ must NOT import from core/ — delegate to services/ instead. Services return TypedDicts and raise ServiceError/ValidationError.
  • New features: Add low-level API in core/client.py → business logic in services/*.py → thin wrappers in mcp/tools/*.py and cli/commands/*.py → tests in tests/services/.
  • Constants: All code-name mappings should be defined in constants.py using the CodeMapper class.

Recent Additions

  • v0.4.6 Batch, Cross-Notebook, Pipelines & Smart Select: Multi-notebook operations, cross-notebook aggregated queries, pipeline workflows, and tag-based notebook discovery. Contributed by @fabianafurtadoff (PR #90).
  • v0.4.6 MCP Tool Consolidation: Consolidated 13 new tools into 4 action-based tools (batch, pipeline, tag, cross_notebook_query), keeping total MCP tools at 35.
  • v0.3.0 Service Layer Refactor: Introduced shared services/ layer with 10+ domain modules, eliminating duplicated logic between CLI and MCP. 576+ unit tests.
  • Skill Commands: nlm skill install/uninstall/list/show for AI assistant integration
  • Verb-First Commands: Alternative command style (nlm install skill, nlm list skills)
  • Interactive Artifact Downloads: download_quiz and download_flashcards with JSON/Markdown/HTML formats
  • Sharing API: notebook_share_status, notebook_share_public, notebook_share_invite for collaboration