NotebookLM Agent Skill

This skill empowers the Agent to use Google NotebookLM as a research, synthesis, and content generation engine.

🧠 Core Capabilities

The Agent can perform the following high-level tasks:

1. Deep Research: Autonomously research a topic on the web and summarize findings.
2. RAG (Retrieval-Augmented Generation): Answer questions based strictly on provided documents (PDF, Drive, URL).
3. Content Studio: Generate “Deep Dive” Audio Overviews (Podcasts), Briefing Docs, FAQs, and Study Quizzes.
4. Knowledge Management: Create/Edit notes and organize sources within notebooks.

🛠️ Tool Reference & Usage Patterns

1. Notebook Operations

Manage the workspace. Always list notebooks first to get valid UUIDs.

# List all notebooks (Required before any operation to find UUIDs)
mcporter call notebooklm.notebook_list --args "{}"

# Create a new project
mcporter call notebooklm.notebook_create --args '{"title": "Project Alpha"}'

# Delete a notebook (Irreversible)
mcporter call notebooklm.notebook_delete --args '{"notebook_id": "UUID", "confirm": true}'

2. Source Management (CRITICAL: Smart Polling)

Adding sources takes time to index. The Agent MUST verify source readiness before querying.

Workflow:

1. Call source_add with wait: true.
2. Check the return value. It must say "ready": true.
3. If "ready": false (timeout), retry adding or wait blindly (status check is not possible).
4. Proceed to query.

# 1. Add Source (URL) - REQUIRED: wait=true
mcporter call notebooklm.source_add --args '{"notebook_id": "UUID", "source_type": "url", "url": "https://example.com", "wait": true, "wait_timeout": 120}'

# 2. Add Source (Text) - REQUIRED: wait=true
mcporter call notebooklm.source_add --args '{"notebook_id": "UUID", "source_type": "text", "text": "Analysis...", "title": "Note", "wait": true}'

# Note: notebook_get does NOT return source readiness status. 
# You MUST rely on source_add(wait=true) returning "ready": true.

Supported Source Types: url (Web/YouTube), text (Pasted), file (Local PDF/MP3/Txt), drive (Google Drive).

3. Deep Research Agent

Delegate complex web research to NotebookLM.

Workflow:

1. research_start -> Returns task_id.
2. research_status -> Loop every 30s until status: "completed".
3. research_import -> Import found sources into the notebook.

# Start Research
mcporter call notebooklm.research_start --args '{"query": "Latest advances in Solid State Batteries", "mode": "deep", "notebook_id": "UUID"}'

# Check Progress (Wait 30s between calls!)
mcporter call notebooklm.research_status --args '{"notebook_id": "UUID", "task_id": "TASK_UUID"}'

# Import Results
mcporter call notebooklm.research_import --args '{"notebook_id": "UUID", "task_id": "TASK_UUID"}'

4. Studio Generation (Podcasts, Quizzes, etc.)

Generate artifacts from sources.

Supported Types:

• audio: Audio Overview (Podcast).
• quiz: Learning Quiz (Multiple Choice).
• faq: Frequently Asked Questions.
• study_guide: Study Guide.
• briefing_doc: Briefing Document.

# Generate Podcast (Deep Dive)
mcporter call notebooklm.studio_create --args '{"notebook_id": "UUID", "artifact_type": "audio", "audio_format": "deep_dive", "confirm": true}'

# Generate Quiz (Great for YouTube videos)
mcporter call notebooklm.studio_create --args '{"notebook_id": "UUID", "artifact_type": "quiz", "confirm": true}'

# Check Status (Wait 30s between calls!)
mcporter call notebooklm.studio_status --args '{"notebook_id": "UUID"}'

# Download Artifact
mcporter call notebooklm.download_artifact --args '{"notebook_id": "UUID", "artifact_type": "audio", "output_path": "./podcast.mp3"}'

5. Querying the Notebook (Critical Field Info)

Once all sources are added and processed, you can ask questions.

NOTEBOOK_ID="...." # Use the ID from the create/list step.

# The response is in the '.answer' field of the result JSON.
npx -y mcporter --config ~/.mcporter_config.json call notebooklm.notebook_query --args "{\"notebook_id\": \"$NOTEBOOK_ID\", \"query\": \"Summarize key findings.\"}"

• IMPORTANT: The JSON response from this call contains the actual answer text in the answer field (e.g., jq -r '.answer'), NOT directly in result or response. Wait patiently for the output.

6. Note Management

Interact with specific notes (atomic thoughts) within a notebook.

# List Notes
mcporter call notebooklm.note --args '{"notebook_id": "UUID", "action": "list"}'

# Create Note
mcporter call notebooklm.note --args '{"notebook_id": "UUID", "action": "create", "content": "My observation...", "title": "Observation 1"}'

🏆 Proven Agent Script (Reference Implementation)

Use this logic for reliable “Deep Research” tasks.

# 1. Define inputs
NOTEBOOK_TITLE="[Agent] Research Task"
SOURCE_URL="https://example.com/source"
QUERY="Summarize key findings."

# 2. Configure mcporter (Safe to run every time)
npx -y mcporter --config ~/.mcporter_config.json config add notebooklm --stdio "bash -c 'export NOTEBOOKLM_COOKIES=\$(cat ~/.nlm/cookies.txt); notebooklm-mcp --transport stdio'"

# 3. Find or Create Notebook
LIST_RESPONSE=$(npx -y mcporter --config ~/.mcporter_config.json call notebooklm.notebook_list --args "{}")
NOTEBOOK_ID=$(echo $LIST_RESPONSE | jq -r --arg title "$NOTEBOOK_TITLE" '.notebooks[] | select(.title == $title) | .id' | head -n 1)

if [ -z "$NOTEBOOK_ID" ]; then
    CREATE_RESPONSE=$(npx -y mcporter --config ~/.mcporter_config.json call notebooklm.notebook_create --args "{\"title\": \"$NOTEBOOK_TITLE\"}")
    NOTEBOOK_ID=$(echo $CREATE_RESPONSE | jq -r '.notebook.id')
    # IMPORTANT: Wait long enough for indexing!
    npx -y mcporter --config ~/.mcporter_config.json call notebooklm.source_add --args "{\"notebook_id\": \"$NOTEBOOK_ID\", \"source_type\": \"url\", \"url\": \"$SOURCE_URL\", \"wait\": true, \"wait_timeout\": 600}"
fi

# 4. Query & Extract .answer field
QUERY_RESPONSE=$(npx -y mcporter --config ~/.mcporter_config.json call notebooklm.notebook_query --args "{\"notebook_id\": \"$NOTEBOOK_ID\", \"query\": \"$QUERY\"}")
FINAL_ANSWER=$(echo $QUERY_RESPONSE | jq -r '.answer')

echo "Final Answer: $FINAL_ANSWER"

• IMPORTANT: The JSON response from this call contains the actual answer text in the answer field (e.g., jq -r '.answer'), NOT directly in result or response. Wait patiently for the output.

🛡️ Best Practices & Failure Prevention

To ensure 100% success rate, the Agent MUST follow these rules:

1. Smart Polling (No Blind Waits):

   • Source Addition: Always use "wait": true. The notebook_get command does not show processing status, so you must rely on the source_add response.
   • Studio/Research: These do support polling. Use studio_status or research_status loops.

2. Rate Limiting:

   • NotebookLM has rate limits. Do not poll faster than once every 10-20 seconds.
   • If you receive a 429 error, sleep for 60 seconds.

3. Source Limits:

   • Keep inputs under 350,000 words or 200MB per source.
   • If a source is too large, split it or summarize it first.

4. Google Drive Access:

   • Ensure the cookies.txt account has permissions to the Drive file.
   • For shared files, make sure they are “Anyone with link can view” or shared directly.

⚙️ Configuration (User Setup)

For the Agent to work, notebooklm-mcp must be authenticated.

Robust Auth Setup (Recommended):

1. Save browser cookies to ~/.nlm/cookies.txt.
2. Register with auth injection:

Linux / macOS:

npx -y mcporter --config ~/.mcporter_config.json config add notebooklm --stdio \
  "bash -c 'export NOTEBOOKLM_COOKIES=\$(cat ~/.nlm/cookies.txt); notebooklm-mcp --transport stdio'"

Windows (PowerShell):

npx -y mcporter --config "$HOME\.mcporter_config.json" config add notebooklm --stdio \
  "powershell -c \"`$env:NOTEBOOKLM_COOKIES = Get-Content '$HOME\.nlm\cookies.txt' -Raw; notebooklm-mcp --transport stdio\""