MCP server bridging Google NotebookLM with Claude Code/Desktop, Cursor, Cline, Continue. Cross-platform auth (macOS/Windows/Linux), auto-refresh, 27 tools.
The server provides 14 tools with clear verb-based naming (list_, create_, add_, ask_, generate_, get_). All tools have descriptions and input schemas are visible and properly typed. However, several critical gaps reduce the score: (1) Many parameter descriptions are minimal (e.g., 'The ID of the notebook' repeated verbatim across tools without explaining when to use which tool); (2) Output schemas are not explicitly documented, responses are inferred from code but not formally specified for the LLM; (3) Error handling is absent, no guidance on what to do if a tool fails, how to recover, or what constitutes a retryable vs. fatal error; (4) No pagination support documented for list_notebooks despite potentially returning large result sets; (5) Tool compositions don't clearly chain (e.g., create_notebook returns an ID but it's unclear if that ID is immediately usable by add_source_* tools); (6) Several tools like generate_* accept optional 'instructions' but don't document what constraints the NotebookLM API places on them. The server is functional and the basic naming convention is sound, but falls short of production-grade clarity for LLM-driven invocation.
Add a Google Drive file as a source. file_id is the Drive ID; default mime_type = Google Doc. Common types: application/pdf, application/vnd.google-apps.spreadsheet, application/vnd.google-apps.presentation.
Add a local file (PDF, DOCX, TXT, MD, etc.) as a source. file_path is absolute or ~-relative.
Add raw text as a source to a notebook.
Add a website URL as a source to a notebook.
Ask a question based on the sources in a specific notebook.
Create a new notebook with the given title.
Output schemas not formally documented. All 14 tools return structured JSON responses, but the LLM has no explicit declaration of what fields to expect. This forces LLMs to infer structure from examples or error messages, increasing hallucination risk.
No error handling or recovery guidance. Tools lack descriptions of what errors can occur (auth failures, rate limits, invalid params) and what the LLM should do next (retry, ask user, fallback). The code does retry on auth errors internally, but this is invisible to the agent.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
Generate an Audio Overview (Deep Dive podcast) in NotebookLM.
Generate an Infographic in NotebookLM. orientation: LANDSCAPE | PORTRAIT | SQUARE. detail_level: CONCISE | STANDARD | DETAILED.
Generate an interactive Mind Map in NotebookLM (saved as a note).
Generate a Quiz in NotebookLM.
Generate a Slide Deck (PowerPoint style) in NotebookLM.
Generate a Video Overview artifact in NotebookLM.
Get the summary and key insights of a notebook.
List all notebooks in your NotebookLM account.
Parameter descriptions are generic and repetitive. Many describe only the parameter itself without explaining WHEN to use the tool or HOW the parameter affects behavior. E.g., 'The ID of the notebook' is repeated identically across 11 tools without clarifying the tool's unique purpose.
No pagination or result-limiting guidance for list_notebooks. The tool returns all notebooks but doesn't document whether results are paginated, what the max count is, or if large account sizes cause performance issues.
generate_* tools lack documentation of async behavior and polling patterns. Descriptions don't explain whether generation is synchronous or async, how long it takes, when results are available, or how to retrieve them. No mention of polling endpoints or webhooks.
Tool composition not explicit. Tools that produce IDs (create_notebook) are used by downstream tools (add_source_*), but descriptions don't state this dependency or guarantee that a freshly created notebook_id is immediately usable.
Unconstrained optional parameters. 'instructions' parameters in generate_* tools have no enum values, length limits, or format guidance. LLMs may pass arbitrary text that the NotebookLM API rejects with unhelpful errors.
Undocumented API constraints. add_source_* tools don't specify file size limits, supported MIME types (beyond examples in add_source_drive), URL length limits, or text length limits. LLMs may pass oversized inputs that fail.