notebooklm（NotebookLM 笔记研究）

NotebookLM

NotebookLM = external RAG engine that offloads reading from your context window. Uses the notebooklm-mcp-cli MCP server (PyPI, v0.5.0+) to create notebooks, manage sources, generate content, and query with grounded AI responses. Supports batch operations across notebooks, pipelines, and multilingual content generation.

Disclaimer: Uses internal undocumented Google APIs via browser authentication. Sessions last ~20 minutes. API may change without notice.

What's New (March 2026)

• Cinematic Video Overviews (Mar 4) — fully animated narrated videos powered by Gemini 3 + Veo 3. Google AI Ultra only, 20/day limit, English only.
• 4 Audio Formats — Brief, Critique, Debate, Deep Dive (was single podcast style)
• 8x Source Capacity — 8x more source material per conversation, 6x extended memory
• Per-Slide Editing — studio_revise edits individual slides without regenerating the full deck
• 3-Panel UI — Sources / Chat / Studio layout on notebooklm.google.com
• Research timeout (v0.5.0) — research_import now configurable via --timeout / timeout param (default 300s, was 120s)
• Deep research errors (v0.5.1) — RPCError class with error codes, auto-retry on transient API failures

Prerequisites

1. Install: uv tool install notebooklm-mcp-cli (or pip install notebooklm-mcp-cli)
2. Authenticate: nlm login (opens browser, session ~20 min)
3. Configure MCP: nlm setup add claude-code (auto-configures .mcp.json) or nlm setup add all for multi-tool setup
4. Verify: nlm login --check to confirm active session
5. Upgrade: uv tool upgrade notebooklm-mcp-cli — restart MCP server after upgrade

CRITICAL: Task Management is MANDATORY (CC 2.1.16)

BEFORE doing ANYTHING else, create tasks to track progress:

# 1. Create main task IMMEDIATELY
TaskCreate(
  subject="NotebookLM: {operation}",
  description="Managing notebooks, sources, and content generation",
  activeForm="Managing NotebookLM resources"
)

# 2. Create subtasks for the notebook workflow
TaskCreate(subject="Notebook setup", activeForm="Creating/configuring notebook")
TaskCreate(subject="Source management", activeForm="Adding sources to notebook")
TaskCreate(subject="Content generation", activeForm="Generating studio content")

# 3. Set dependencies for sequential steps
TaskUpdate(taskId="3", addBlockedBy=["2"])
TaskUpdate(taskId="4", addBlockedBy=["3"])

# 4. Before starting each task, verify it's unblocked
task = TaskGet(taskId="2")  # Verify blockedBy is empty

# 5. Update status as you progress
TaskUpdate(taskId="2", status="in_progress")  # When starting
TaskUpdate(taskId="2", status="completed")    # When done

Decision Tree — Which Rule to Read

What are you trying to do?
│
├── Create / manage notebooks
│   ├── List / get / rename ──────► notebook_list, notebook_get, notebook_rename
│   ├── Create new notebook ──────► notebook_create
│   └── Delete notebook ──────────► notebook_delete (irreversible!)
│
├── Add sources to a notebook
│   ├── URL / YouTube ────────────► source_add(type=url)
│   ├── Plain text ───────────────► source_add(type=text)
│   ├── Local file ───────────────► source_add(type=file)
│   ├── Google Drive ─────────────► source_add(type=drive)
│   ├── Rename a source ──────────► source_rename
│   └── Manage sources ──────────► rules/setup-quickstart.md
│
├── Query a notebook (AI chat)
│   ├── Ask questions ────────────► notebook_query
│   └── Configure chat style ────► chat_configure
│
├── Generate studio content
│   ├── 10 artifact types ───────► rules/workflow-studio-content.md
│   ├── Revise slides ───────────► studio_revise (creates new deck)
│   └── Export to Docs/Sheets ──► export_artifact
│
├── Research & discovery
│   └── Web/Drive research ──────► rules/workflow-research-discovery.md
│
├── Notes (capture insights)
│   └── Create/list/update/delete ► note (unified tool)
│
├── Sharing & collaboration
│   └── Public links / invites / batch ► rules/workflow-sharing-collaboration.md
│
├── Batch & cross-notebook
│   ├── Query across notebooks ────► cross_notebook_query
│   ├── Bulk operations ───────────► batch (query, add-source, create, studio)
│   └── Multi-step pipelines ──────► rules/workflow-batch-pipelines.md
│
├── Organization
│   └── Tag notebooks ─────────────► tags
│
└── Workflow patterns
    ├── Second brain ─────────────► rules/workflow-second-brain.md
    ├── Research offload ─────────► rules/workflow-research-offload.md
    └── Knowledge base ──────────► rules/workflow-knowledge-base.md

Quick Reference

Total: 9 rules across 5 categories

MCP Tools by API Group

Total: 37 tools across 13 groups (v0.5.0+)

Key Decisions

Example

# 1. Create a notebook for your project
notebook_create(title="Auth Refactor Research")

# 2. Add sources (docs, articles, existing code analysis)
source_add(notebook_id="...", type="url", url="https://oauth.net/2.1/")
source_add(notebook_id="...", type="text", content="Our current auth uses...")
source_add(notebook_id="...", type="file", path="/docs/auth-design.md")

# 3. Query with grounded AI responses
notebook_query(notebook_id="...", query="What are the key differences between OAuth 2.0 and 2.1?")

# 4. Generate a deep dive audio overview (supports language param)
studio_create(notebook_id="...", artifact_type="audio", audio_format="deep_dive", language="he", confirm=True)
studio_status(notebook_id="...")  # Poll until complete

# 5. Generate a cinematic video overview (Ultra only, English)
studio_create(notebook_id="...", artifact_type="cinematic_video", confirm=True)
studio_status(notebook_id="...")  # Poll — takes 3-8 minutes

# 6. Capture insights as notes
note(notebook_id="...", action="create", content="Key takeaway: PKCE is mandatory in 2.1")

Common Mistakes

• Forgetting auth expiry — Sessions last ~20 min. Always check with nlm login --check before long workflows. Re-auth with nlm login.
• One giant notebook — Split by project/topic. One notebook with 50 sources degrades retrieval quality.
• Huge single sources — Split documents >50K characters into logical sections for better chunking and retrieval.
• Not polling studio_status — Studio content generation takes 2-5 minutes. Poll studio_status instead of assuming instant results.
• Ignoring source types — Use type=url for web pages (auto-extracts), type=file for local files. Using type=text for a URL gives you the URL string, not the page content.
• Deleting notebooks without checking — notebook_delete is irreversible. List contents with source_list_drive and note(action=list) first.
• Skipping research_import — research_start discovers content but does not add it. Use research_import to actually add findings as sources.
• Raw queries on empty notebooks — notebook_query returns poor results with no sources. Add sources before querying.
• Ignoring language param — studio_create supports BCP-47 language codes (e.g., he, ar, ja). Defaults to English if omitted.
• Batch without purpose — batch and cross_notebook_query are powerful but add latency. Use for multi-project synthesis, not single-notebook tasks.

Related Skills

• ork:mcp-patterns — MCP server building, security, and composition patterns
• ork:web-research-workflow — Web research strategies and source evaluation
• ork:memory — Memory fabric for cross-session knowledge persistence
• ork:security-patterns — Input sanitization and layered security