NotebookLM Integration for AI Coding Agents¶

Bridge between Claude Code and Google NotebookLM via reverse-engineered CLI (notebooklm-py). Heavy analytics (30+ documents, web research, cross-references) runs free on Google's Gemini RAG infrastructure. Claude spends tokens only on orchestration and final editing.

Architecture¶

User request
    |
    v
Claude Code (orchestrator, token-metered)
    |
    v  CLI calls
notebooklm-py (bridge, reverse-engineered API)
    |
    v
Google NotebookLM (Gemini RAG, free tier: 50 sources/notebook)
    |
    v  results
Claude Code (final polish, minimal tokens)

Setup¶

# Install CLI bridge
# Repo: https://github.com/teng-lin/notebooklm-py
# Follow README for installation + Google auth

# Install Claude Code skill
notebooklm skill install
notebooklm skill status    # verify

Skill installs to ~/.claude/skills/notebooklm/SKILL.md (personal) and ~/.agents/skills/notebooklm/ (cross-agent compatible with Codex, Gemini CLI).

Four Workflows¶

A. Research Without Token Spend¶

Offload document analysis to NotebookLM. Claude orchestrates, Google processes.

# 1. Create notebook
notebooklm create "My Research Project"

# 2. Add sources (up to 50 free, 300 Pro)
notebooklm source add \
    "./transcript-1.md" \
    "https://example.com/article" \
    "./report.pdf"

# 3. Query across all sources (free, Gemini RAG)
notebooklm ask "what are the three most important themes across all sources?"

# 4. Generate artifacts
notebooklm generate slide-deck
notebooklm generate flashcards --quantity more
notebooklm generate mind-map
notebooklm generate data-table "compare key concepts"
notebooklm generate audio "make it engaging" --wait

Token math: analytical work on Google infrastructure. Claude tokens reserved for orchestration + final editing only. $20/month plan stretches to $200/month capability.

B. Expert Agent from Web Research (DBS Framework)¶

Use NotebookLM Deep Research for autonomous web crawling, then structure into a Claude Code skill.

1. Run Deep Research in NotebookLM browser (source type: "web", specific query)
2. Structure results using DBS framework:
3. Direction - step-by-step logic, decision trees, error handling -> becomes SKILL.md core
4. Blueprints - static references: templates, tone guidelines, classification rules -> companion files
5. Solutions - deterministic code tasks: API calls, formatting, calculations -> scripts
6. Feed to /skill-creator in Claude Code -> auto-generates full skill package
7. Test and iterate

C. Cross-Session Memory via Master Brain¶

# End of session: extract insights
# /wrap-up skill extracts: corrections, successful patterns, open questions, decisions

# Push to dedicated NotebookLM notebook
notebooklm use master-brain-notebook-id \
    "./session-summary-2026-04-06.md"

# Add to CLAUDE.md:
# "Before answering architecture questions, query Master Brain via NotebookLM CLI"

Over weeks, Master Brain accumulates hundreds of session summaries. NotebookLM indexes everything with semantic connections. Claude retrieves exactly the context needed without loading hundreds of documents into its context window.

D. Visual Knowledge via Obsidian¶

Run Claude Code from Obsidian vault root. All generated files appear in Obsidian's graph view.

cd ~/Documents/MyVault
claude

CLAUDE.md in vault root specifies: folder structure, mandatory metadata (dates, tags, source links), cross-reference rules ([[double brackets]]), formatting standards.

Custom skills: /research <topic>, /daily (daily summary with cross-refs), /wrap-up (session memory to vault).

CLI Reference¶

Gotchas¶

• Unofficial API - no stability guarantee. notebooklm-py reverse-engineers Google's internal protocols. Google can break it at any time by changing their backend. No SLA, maintained by one developer (Teng Ling). Treat as power-user tool, not production infrastructure
• storage_state.json contains live Google session cookies. Anyone with this file gets full access to your NotebookLM data. Never commit to git. Treat as a password. Add to .gitignore immediately
• Cookies expire periodically. When commands fail with auth errors, run notebooklm login (30 seconds). No way to get permanent tokens
• GDPR implications. User-tier Claude and NotebookLM process/store data in the US. If you work with regulated data (EU/UK), corporate API offers regional processing but consumer tier does not
• Anthropic ToS compliance. Do not use this to circumvent token limits through unofficial wrappers. Ensure usage matches your plan

See Also¶

• context engineering - context management strategies
• agent memory - memory patterns for AI agents
• rag pipeline - RAG architecture that NotebookLM uses internally
• token optimization - reducing token consumption
• managed agents - Anthropic's official hosted agent platform