Accessing NotebookLM via MCP

Plan → Validate → Execute (required pattern)

Plan

• Restate the user goal in one sentence.
• Pick the minimal set of MCP calls needed.

Validate

• Verify notebook_id exists (or create it).
• Verify sources exist / are imported before querying.
• For Drive sync: run source_list_drive first.
• For destructive/studio actions: obtain explicit user approval before calling with confirm=True.

Execute

• Perform the MCP calls.
• Report the output IDs/URLs and what changed.

Instructions

1) Authentication & session recovery

Use this flow whenever calls fail due to auth or right after running CLI auth.

1. Call refresh_auth.
2. If unsuccessful:
   • Prefer running the CLI in terminal: notebooklm-mcp-auth (see scripts/).
   • Then call refresh_auth again to pick up new tokens/cookies.
3. If CLI auth fails and the user explicitly provides cookie headers, use save_auth_tokens as a fallback.

Rules

• Do not proceed with notebook/source operations if auth is not valid.
• When unsure how to run a helper script, run it with --help.

2) Notebook selection & management

Select a notebook

• If the user provides a notebook_id, use it.
• Else:
  1. Call notebook_list (default max_results is fine).
  2. Match by title (case-insensitive substring).
  3. If no clear match, create a notebook with notebook_create using a descriptive title.

Common operations

• List: notebook_list(max_results=...)
• Create: notebook_create(title=...)
• Details + sources: notebook_get(notebook_id=...)
• AI summary + suggested topics: notebook_describe(notebook_id=...)
• Rename: notebook_rename(notebook_id=..., new_title=...)

Delete (irreversible)

• Only call notebook_delete(notebook_id=..., confirm=True) after explicit user approval that includes “delete” intent.
• Before deleting, fetch context with notebook_get and summarize what will be removed.

3) Adding sources (URL, text, Drive)

Add a URL or YouTube transcript

• Use notebook_add_url(notebook_id=..., url=...).
• After adding, call notebook_get to confirm the source appears.

Add pasted text

• Use notebook_add_text(notebook_id=..., text=..., title=optional).

Add Google Drive documents

• Use notebook_add_drive(notebook_id=..., document_id=..., title=..., doc_type=doc|slides|sheets|pdf).
• Extract document_id from the Drive URL (the long ID in /d/<ID>/ or similar).

Validation tips

• After adding sources, call notebook_get to obtain source IDs.
• For quick export of indexed text, prefer source_get_content over notebook_query.

4) Source inspection, export, and deletion

Describe (AI summary + keyword chips)

• Use source_describe(source_id=...) to generate a short summary and keywords.

Export raw indexed content (fast)

• Use source_get_content(source_id=...) to retrieve the original indexed text and metadata.

Delete a source (irreversible)

• Only call source_delete(source_id=..., confirm=True) after explicit user approval.
• Confirm the correct source by title/type via notebook_get before deleting.

5) Asking questions about existing sources (NOT web search)

Use notebook_query only when the needed sources are already in the notebook.

Steps

1. Ensure sources exist:
   • notebook_get and capture relevant source_ids (or query all).
2. Ask:
   • notebook_query(notebook_id=..., query=..., source_ids=[...optional...], conversation_id=...optional..., timeout=...optional...)

Follow-ups

• Reuse conversation_id for iterative questioning.

If responses time out

• Narrow to specific source_ids.
• Increase timeout if the environment allows.

6) Research workflow (find NEW sources via web or Drive)

Use this lane when the user wants discovery, “find sources,” “deep research,” “search web,” or “search Drive.”

Correct sequence

1. research_start(query=..., source=web|drive, mode=fast|deep, notebook_id=optional, title=optional)
2. research_status(notebook_id=..., task_id=optional, poll_interval=30, max_wait=300, compact=True)
3. When status is completed:
   • research_import(notebook_id=..., task_id=..., source_indices=optional)

Heuristics

• Use mode=fast for quick results and iteration.
• Use mode=deep for comprehensive web-only research.
• After import, call notebook_describe to generate a notebook-level summary and suggested topics.

7) Drive source freshness & sync

When the user asks to “sync Drive docs” or “update sources”:

1. Call source_list_drive(notebook_id=...).
2. Identify stale/out-of-date sources.
3. Ask for explicit approval to sync.
4. Call source_sync_drive(source_ids=[...], confirm=True).
5. Confirm updates by re-running source_list_drive or notebook_get.

8) Chat configuration (goal + length)

Use chat_configure to steer style:

• goal=default|learning_guide|custom
• If goal=custom, provide custom_prompt (<= 10000 chars).
• response_length=default|longer|shorter

Pattern

• Configure once per notebook session unless the user changes requirements.

9) Studio artifact generation (requires explicit approval)

These calls require confirm=True only after the user explicitly approves generating the artifact.

Available artifacts

• Audio overview: audio_overview_create(...)
• Video overview: video_overview_create(...)
• Infographic: infographic_create(...)
• Slide deck: slide_deck_create(...)
• Report: report_create(...) (use “Create Your Own” only with a custom prompt)
• Flashcards: flashcards_create(...)
• Quiz: quiz_create(...)
• Data table: data_table_create(...)
• Mind map: mind_map_create(...)

Execution pattern

1. Validate sources exist (via notebook_get), optionally limit source_ids.
2. Obtain explicit user approval.
3. Call the create function with confirm=True.
4. Poll studio_status(notebook_id=...) and return any URLs or artifact IDs.

Delete a studio artifact (irreversible)

• Only after explicit user approval:
  • studio_delete(notebook_id=..., artifact_id=..., confirm=True)

10) Minimal call templates (copy/paste patterns)

Auth refresh

refresh_auth()