NotebookLM via MCP

This skill uses the @m4ykeldev/notebooklm-mcp server, which exposes NotebookLM as 27 MCP tools. NotebookLM has no public API; the server talks to the same internal RPC endpoint the web app uses, authenticated with the user's own Google session cookies.

Setup

Check first: if tools such as notebook_list are already available, skip this section.

1. Authenticate once (opens Chrome, waits for the Google login, stores cookies in ~/.notebooklm-mcp/auth.json):

   npx -y @m4ykeldev/notebooklm-mcp auth

   Headless / CI: set NOTEBOOKLM_COOKIES instead. Fallbacks: auth --manual (paste cookies) or auth --file tokens.json.

2. Register the server with the client.

   Claude Code:

   claude mcp add --transport stdio notebooklm -- npx -y @m4ykeldev/notebooklm-mcp serve

   Any JSON-configured client (Cursor, VS Code, Claude Desktop, ...):

   {
     "mcpServers": {
       "notebooklm": {
         "command": "npx",
         "args": ["-y", "@m4ykeldev/notebooklm-mcp", "serve"]
       }
     }
   }

   Long answers on big notebooks: append "--query-timeout", "180000" to args.

The authentication step needs the user at a browser. Ask them to run it; do not try to log in on their behalf.

Core workflow

1. Find or create the notebook. notebook_list returns IDs, titles and sources, up to max_results (default 100); if the count reaches it, call again with a higher value. Match by title before creating a duplicate. notebook_create returns the new ID.
2. Add sources.
   • source_add_url — web pages and YouTube links.
   • source_add_text — pass content inline, or file_path for a local file. file_path must sit under the server's working directory or the OS temp dir; for anything else, read the file and pass content.
   • source_add_drive — needs the Drive file_id, a title and the MIME doc_type (e.g. application/vnd.google-apps.document).
   • Each of these returns the new source's id. Sources take a moment to process; check with notebook_get before querying one you just added.
   • If source_add_url says NotebookLM could not fetch the page, check notebook_get before retrying: the source is sometimes listed anyway.
3. Ask questions. notebook_query answers only from the notebook's sources, using all of them by default. Restrict with source_ids when the user cares about specific documents. For follow-ups, pass back the returned conversation_id so the follow-up keeps context (the server remembers the conversation while it keeps running).
4. Produce output with the Studio tools (below), or summarise with notebook_summarize.

Prefer notebook_query over reading raw sources. Use source_get_content only when the user needs the verbatim text of a source.

Research

Use when the user wants NotebookLM to find sources for them.

1. research_start with query, source (web default, or drive) and mode (fast default, or deep for a longer, more thorough pass). Returns a task_id.
2. Poll research_status until the status leaves in_progress. Deep research can take minutes; poll at a relaxed pace and tell the user it is still running.
3. Show the user what was found, then research_import with source_indices for the ones they want. Indices are positions in the sources list returned by research_status. Only web results (those with a URL) can be imported: omitting source_indices imports every web result and skips Drive ones, while an index pointing at a Drive result is an error.
4. Read the result: "Imported X of Y research sources". NotebookLM fetches each URL and silently skips the ones it can't reach, so X can be lower than Y. If none could be fetched the tool returns an error and the task shows import_failed; retrying usually works, or pick other results.

Studio content

Every generator needs confirm: true. These consume the user's NotebookLM quota and take minutes, so confirm the type and options with the user before calling. confirm is a required boolean: with confirm: false the tool returns pending_confirmation and does nothing; omitting it is a validation error.

All accept source_ids to limit which sources are used (omit for all). language is a BCP-47 code such as en or es; without it NotebookLM uses the account's language.

After starting one, poll studio_status for the notebook and find the artifact by the returned artifact_id: it lists older artifacts too. Each artifact reports pending, generating, completed or failed. Audio, video, infographics and slide decks (as PDF) get a download_url once completed. Give the user the URL; the download needs their Google session, so it opens in their browser. Reports, flashcards and quizzes have no file: point the user to the notebook in NotebookLM to read them.

Housekeeping

• notebook_rename, notebook_get (sources and their IDs), source_get (one source's metadata).
• source_sync_drive re-pulls the latest version of Drive sources.
• notebook_delete and source_delete are permanent and need confirm: true. Name the exact notebook or source and get explicit approval from the user first.

When things fail

• Auth errors (expired session, 401/403): call auth_refresh once. The server also tries to restore the session in the background. If it still fails, or the error says the session is signed out (redirected to Google sign-in, no CSRF token), ask the user to run npx -y @m4ykeldev/notebooklm-mcp auth again, then retry.
• Timeouts on notebook_query: suggest raising --query-timeout, or narrow the question with source_ids.
• Unexpected or empty responses: Google changes NotebookLM's internal API without notice. Report the error to the user rather than retrying in a loop, and suggest updating the package (npx -y @m4ykeldev/notebooklm-mcp@latest serve) or opening an issue on the GitHub repo.
• auth_save_tokens is a manual fallback for cookies. Use it only when the user hands you cookies explicitly, and never echo cookie values back in chat.