NotebookLM MCP & Automation Integration Comparison

This report evaluates different approaches to integrating Google NotebookLM with AI agents (such as Claude Code, Cursor, or Codex) and command-line workflows. It compares the current browser-automation-based skill against two leading open-source alternatives: PleasePrompto/notebooklm-mcp and jacob-bd/notebooklm-mcp-cli.

1. Context & Problem Statement

The current /notebooklm skill used in this workspace is based on Browser Automation (driving a headless Chromium instance via Patchright/Playwright). In practice, this approach suffers from severe limitations:

High Latency (Slow Responses): Each question requires launching a new browser context, navigating to the NotebookLM page, waiting for the chat history to stabilize (5+ seconds), typing the question character-by-character, waiting for the streaming response, and shutting down. This takes 15 to 45 seconds per query.
Selector Fragility (UI Drift): Google frequently updates the NotebookLM frontend DOM structure and class names. Because the scraper relies on specific CSS selectors (like .to-user-container .message-text-content), any UI update breaks the scraper, leading to timeouts or empty responses.
No Source/Citation Extraction: The current script only extracts the plain text of the response. It does not extract the grounded sources, document titles, page numbers, or exact quotes, which defeats the core value proposition of NotebookLM.
Stuck/Repeating Responses: If the browser session fails to detect the end of a stream or reads a cached DOM state, it can return stale responses or get stuck in a loop.

2. Technical Comparison Matrix

This matrix compares the current implementation against the two open-source alternatives.

Name (Link)	Stars	Latest Commit Date	Setup Overhead	Pros & Cons	License & Cost	Key Core Features
Current `/notebooklm` Skill (Baseline)	N/A	Local	Medium (Auto-venv via `run.py`)	+ Stateless, isolated - Extremely slow (15-45s) - No source extraction - Highly fragile to UI updates	MIT / Free	Basic Q&A via headless browser scraping
PleasePrompto/notebooklm-mcp	Active	2026-05-01	Medium (Installed via `npx`)	+ Full MCP server protocol + Cross-client session sharing - Still uses Patchright browser automation under the hood - High CPU/RAM footprint	MIT / Free	Q&A, Notebook & source management, Audio Overview generation
jacob-bd/notebooklm-mcp-cli	Active	2026-06-25	Medium (Installed via `pipx` / `pip`)	+ Extremely Fast (Direct HTTP requests) + Rich Source Extraction (JSON metadata) + No UI selector fragility - Risk of "RPC drift" if Google changes internal endpoints	MIT / Free	Q&A with precise citations, full notebook/source CRUD, Audio Overview, Study Guide generation

3. Deep-Dive Analysis of Alternatives

A. PleasePrompto/notebooklm-mcp (Browser-Driven MCP)

How it works: It uses TypeScript/Node.js to run a Model Context Protocol (MCP) server. It drives a real Chromium browser using Patchright (a Playwright fork designed to bypass bot detection) to automate the NotebookLM web interface.
Why it fails to solve the current issues: Since it uses the same underlying browser-automation architecture as our current skill, it does not solve the speed or selector fragility problems. When Google updates its UI, the selectors in src/notebooklm/selectors.ts must be manually updated by the maintainers or the server will fail.

B. jacob-bd/notebooklm-mcp-cli (API/RPC-Driven CLI & MCP)

How it works: This Python-based tool bypasses the browser UI entirely for queries. It uses a one-time login command (nlm login) to extract session cookies using the Chrome DevTools Protocol (CDP). Once authenticated, it communicates directly with Google's internal batchexecute RPC endpoints by sending raw POST payloads.
Why it is the superior approach:
Speed: Because it makes direct HTTP requests without the overhead of rendering a browser or simulating human typing, queries return in 1 to 3 seconds (up to 10x faster).
Data Richness: The raw RPC responses contain structured JSON metadata for every citation. The tool can extract the exact source document title, page numbers, and the specific quote used to ground the answer, formatting them beautifully into markdown footnotes.
Stability: It is immune to CSS class changes or DOM layout modifications.
RPC Drift Mitigation: To handle Google changing internal RPC method IDs, the tool supports a JSON environment variable NOTEBOOKLM_RPC_OVERRIDES. If Google updates a method ID, the user can override it instantly without waiting for a package update.

4. Recommended Winner & Action Plan

Recommended Winner: `jacob-bd/notebooklm-mcp-cli`

This tool directly addresses all the pain points of the current system: 1. It eliminates the 30-second browser startup delay. 2. It provides complete, structured source and citation extraction. 3. It is much more stable than DOM scraping.

Implementation & Transition Plan

Install notebooklm-mcp-cli: bash pipx install notebooklm-mcp-cli
Authenticate: bash nlm login This opens a browser window for a one-time Google login and automatically captures the session cookies.
Configure as an MCP Server: Add it to the Claude / Antigravity configuration (e.g., mcp.json or through the CLI): json { "mcpServers": { "notebooklm": { "command": "nlm", "args": ["mcp"] } } }
Deprecate the Browser-Based Skill: Remove or archive the /Users/keira/.gemini/config/skills/notebooklm directory to prevent the agent from falling back to the slow browser-based automation.