NotebookLM MCP Structured MCP Server

NotebookLM MCP Structured

Enhanced MCP server for NotebookLM with client-side prompt structuring for source fidelity.

Requirements: This MCP server is designed to work with Claude Desktop. It requires Claude Desktop to be installed and configured to use MCP servers.

This is a modified version of notebooklm-mcp that adds comprehensive structuring instructions to guide Claude in crafting prompts that enforce document fidelity for professional use cases (legal analysis, research, fact-checking).

Key Features

Client-Side Prompt Structuring

The MCP tool description includes comprehensive guidelines that instruct Claude on how to structure questions before sending them to NotebookLM. This ensures:

• Source fidelity: Responses come ONLY from uploaded documents
• Citation requirements: Every claim includes source attribution
• Missing information handling: Explicit declaration when data is unavailable
• Multi-language support: Works naturally with any language Claude supports
• Question type adaptation: Different structures for comparisons, lists, analyses, explanations, and extractions

How It Works:

1. User asks a simple question in any language
2. Claude reads the structuring guidelines from the tool description
3. Claude transforms the question into a well-structured prompt
4. NotebookLM receives the structured prompt and responds accordingly
5. Claude is instructed to present the response faithfully without adding external knowledge

Why This Matters:

NotebookLM already provides source fidelity by design (Gemini grounded on documents). The real problem this fork solves is different: preventing Claude from "improving" NotebookLM's responses with external knowledge when presenting them to the user.

Design Intent (what the fork aims to achieve):

┌─────────────────────────────────────────────────────┐
│ Without structuring (original MCP):                 │
│ • NotebookLM: "Document states X [Source: doc.pdf]"│
│ • Risk: Claude may add external knowledge          │
│   "Document states X. Also, based on my knowledge, │
│   Y is important to consider..."                   │
│   └─ External knowledge added! ─┘                  │
│                                                     │
│ With structuring (this fork):                       │
│ • NotebookLM: "Document states X [Source: doc.pdf]"│
│ • Claude reads Response Handling instruction        │
│ • Goal: Claude presents faithfully                  │
│   "Document states X [Source]"                     │
│   └─ Faithful presentation, no additions ─┘        │
└─────────────────────────────────────────────────────┘

The structuring guidelines include two critical instruction phases:

1. Pre-send: Transform questions with explicit constraints (but preserve original wording)
2. Post-receive: Instruct Claude to present responses faithfully WITHOUT external knowledge

This dual-phase approach is designed to maintain document fidelity throughout the workflow.

Verification & Transparency

How to verify the workflow:

Since NotebookLM saves chat history in your notebooks, you can verify the entire process:

1. Ask a question through Claude using this MCP
2. Open your notebook on the NotebookLM web interface (https://notebooklm.google.com)
3. View the saved chat to see:
   • The structured prompt that Claude sent (via the MCP)
   • The original NotebookLM response with all internal reference links

What you can verify:

• That structuring was applied correctly to your question
• The raw NotebookLM response before Claude presents it
• How Claude interpreted the Response Handling instructions
• Which language was used for structured prompts (useful for multilingual testing)

This transparency mechanism lets you empirically verify the client-side structuring approach and understand each phase of the workflow.

Example Transformation:

Simple question:

What are the main findings in the research papers?

Claude structures it as:

RESPONSE INSTRUCTIONS

TASK: What are the main findings in the research papers?

OPERATIONAL CONSTRAINTS
- Use ONLY information explicitly present in uploaded documents
- DO NOT add external knowledge or interpretations
- If information is not present, state: "[NOT FOUND IN DOCUMENTS]"

REQUIRED OUTPUT FORMAT
For each finding:
- FINDING: [description]
- SOURCE: [document name/section]
- QUOTE: "direct quote supporting the finding"

CITATIONS
- Every claim MUST include source
- Use direct quotes where possible

HANDLING MISSING INFORMATION
- If information is missing, declare it explicitly

BEGIN STRUCTURED RESPONSE

Critical Formatting Rule:

• NO decorative lines (no === or ---) as they cause NotebookLM timeouts

Language Support

Multilingual by design - The fork works with multiple languages without requiring server-side configuration.

**Ho