NotebookLM MCP · Gemini Notebook · Claude · Updated August 9, 2026
How to Connect NotebookLM to Claude Desktop with MCP
This step-by-step NotebookLM MCP guide covers the current Gemini Notebook bridge for Claude Desktop and Claude Code: one-click extension setup, CLI-managed setup, token economics, security, use cases, and troubleshooting.
Unofficial community connectorNotebookLM → Gemini Notebook≈10 min Desktop setup.mcpb + CLI pathsWindows, macOS + Linux
Name update: Google renamed NotebookLM to Gemini Notebook in July 2026. The featured GitHub repository is now gemini-notebook-mcp-cli, while the PyPI package notebooklm-mcp-cli and executable notebooklm-mcp remain for compatibility. This page keeps “NotebookLM MCP” in the title because users and package names still use that term.
Direct answer
How do you connect NotebookLM / Gemini Notebook to Claude?
Claude Desktop: use the current .mcpb extension for the simplest install, or install notebooklm-mcp-cli and run nlm setup add claude-desktop for CLI-managed configuration. Claude Code: install the same package, authenticate with nlm login, and run nlm setup add claude-code.
Choose .mcpb for the easiest Desktop path or CLI setup for terminal-based management.
Fully quit Claude Desktop before a CLI-managed Desktop configuration change.
Authenticate with the Google account that owns the notebooks.
Verify with a read-only notebook-list prompt before allowing write actions.
Before installing: the featured community package uses undocumented Gemini Notebook interfaces and browser-session authentication. Google can change those interfaces or require a new sign-in without notice. Do not treat this as an enterprise-ready connector without independent security review.
Current checkVerified August 9, 2026: GitHub marks v0.9.7 (released August 6) as the latest release. The repository has been renamed to gemini-notebook-mcp-cli. Current setup supports both the one-click .mcpb Desktop extension and nlm setup add claude-desktop; nlm setup add claude-code remains the Claude Code path. Check the latest release before installing ↗
GitHub repositorygemini-notebook-mcp-cli
PyPI packagenotebooklm-mcp-cli
CLI commandnlm
MCP executablenotebooklm-mcp
How this guide was verified: NotebookLM Guide editorial review checked the current GitHub release, README/changelog, Anthropic Claude Desktop and Claude Code documentation, and MCP security guidance. This is an independent guide, not documentation from Google or Anthropic. About NotebookLM Guide →
Local setup
How to install NotebookLM MCP in Claude Desktop
Claude Desktop now has two current setup paths. Use the .mcpb extension for simplicity; use CLI-managed Desktop setup when you want profile-aware configuration and terminal diagnostics. Claude Code uses its own CLI registration command.
Primary path · Claude Desktop
Install the one-click desktop extension
Download the current .mcpb file and install it from Claude Desktop’s Extensions settings. No manual JSON editing.
Advanced option — manage Claude Desktop from the CLI
Current releases support profile-aware Claude Desktop configuration again. Fully quit the Claude Desktop profile you want to change, then run:
nlm setup add claude-desktop
If both regular and Relay AI / 3P profiles are detected, the CLI can prompt for the target profile. Use nlm setup list and nlm doctor to verify the result. For most nontechnical users, the .mcpb extension is still the simpler path.
Path A — Claude Desktop extension
Download the latest .mcpb release
Use the maintained project’s latest release page. Privately distributed extensions do not necessarily update automatically, so return to the release page when this guide or the package changelog announces an update.
Open Settings → Extensions → Advanced settings → Extension Developer → Install Extension. Select the downloaded .mcpb file and approve the installation. Double-clicking the file may open the same installer.
Restart and verify
Fully quit and reopen Claude Desktop. Ask Claude to list notebooks without creating or modifying anything. Complete any Google authentication prompt opened by the connector.
Connection test
Use the NotebookLM MCP tools to list my available notebooks. Return each notebook title, notebook ID, source count, and last-updated date. Do not create, edit, delete, share, or download anything.
Claude Desktop extension does not appear?
Confirm you are on the latest Claude Desktop version, open Settings → Extensions, verify the extension is enabled, and review its logs. Re-download the current .mcpb file if installation failed. The package maintainer removed the old desktop CLI-registration command because it was less reliable.
Path B — Claude Code CLI
Claude Code installed
Google account with Gemini Notebook access
Chrome, Edge, Brave, Arc or Chromium
Terminal or PowerShell access
Install uv
uv installs the package in an isolated tool environment.
Close and reopen the terminal if uv is not immediately recognized.
Install or upgrade the unified package
uv tool install notebooklm-mcp-cli
Already installed?
uv tool upgrade notebooklm-mcp-cli
Authenticate the Google account
The command opens a managed Chromium-family browser profile. Sign into the account that owns the notebooks.
nlm login
nlm login --check
Register the server with Claude Code
nlm setup add claude-code
Then verify the registration and authentication:
nlm setup list
nlm doctor
Already installed an older NotebookLM MCP package?
Remove the separate legacy packages, then install the unified package. Claude Desktop users should reinstall the latest .mcpb file afterward; Claude Code users should run nlm setup add claude-code.
Generate a standard JSON configuration for Cursor, Cline, or another client that accepts MCP JSON. This is not the recommended Claude Desktop path.
nlm setup add json
Maintenance
How to update or remove NotebookLM MCP
Claude Desktop
Extension path: download and install the newer .mcpb release; remove it from Settings → Extensions. CLI-managed path: upgrade the package, fully quit Desktop, then rerun nlm setup add claude-desktop or remove it with nlm setup remove claude-desktop.
Claude Code
Upgrade the package with uv tool upgrade notebooklm-mcp-cli. Remove the client registration with nlm setup remove claude-code, then uninstall the package with uv tool uninstall notebooklm-mcp-cli.
Architecture decision
Should NotebookLM MCP run locally or remotely?
For this community connector, local Claude Desktop or Claude Code is the safest practical default. Remote MCP is possible, but hosting it securely is a separate engineering project.
Recommended for this workflow
Local MCP server
The NotebookLM MCP process and its Google browser profile stay on your computer.
Use .mcpb in Claude Desktop or CLI registration in Claude Code
Supports local file paths and local downloads
No public endpoint or OAuth gateway required
Easier browser-session recovery
Best default for one person
Advanced deployment
Remote custom connector
Claude reaches a publicly hosted MCP endpoint through Anthropic’s cloud infrastructure.
Can appear across web, mobile, Desktop and Claude Code
Requires public HTTPS and caller authentication
Must protect the active Google session
Current community package lacks turnkey multi-user isolation
Do not expose the local server directly
Fastest Desktop pathInstall the current .mcpb desktop extension from the project release page, then complete authentication.Open latest release ↗
Claude Code CLI pathInstall the unified package with uv, run nlm login, and register it with nlm setup add claude-code.Follow the commands ↓
Need browser/mobile access?Keep Claude Code and MCP local, then use Claude Code Remote Control rather than publishing an unsecured MCP endpoint.Read official guidance ↗
Implementation choice
Which Gemini Notebook MCP integration path should you use?
The practical decision is no longer just “which GitHub server?” Choose the installation and governance path that matches where Claude runs and how much control you need.
Comparison of Gemini Notebook and Claude integration paths
Path
Best for
Why choose it
Tradeoff
.mcpb Desktop extension
Most Claude Desktop users
Fastest install; no manual JSON editing; easy to disable or remove from Extensions.
Still relies on a community connector and browser-session authentication.
nlm setup add claude-desktop
Advanced Desktop users
Profile-aware setup, terminal diagnostics, and a consistent CLI-managed workflow.
Requires package installation and careful handling while Desktop is fully closed.
nlm setup add claude-code
Claude Code, automation and developer workflows
Natural fit for scripted pipelines, diagnostics, /mcp, /context, and tool-search controls.
More moving parts than the one-click Desktop extension.
Alternative community MCP implementation
Focused retrieval designs or different client support
May expose a smaller or differently designed tool surface.
Commands and auth behavior differ; do not mix setup instructions across projects.
Official Gemini Notebook Enterprise API
Organization-owned integrations
Documented Google enterprise API path with governance under an enterprise architecture.
Not a drop-in replacement for the consumer community MCP; requires engineering and enterprise configuration.
Selection rule: keep one active Gemini Notebook MCP configuration per client unless you deliberately need multiple implementations. Duplicate server names and legacy configs can make Claude call the wrong tool or make troubleshooting ambiguous.
Decision framework
When is the NotebookLM–Claude bridge actually worth using?
Use MCP when the handoff between grounded evidence and Claude’s reasoning has become the bottleneck—not merely because a connector exists.
Without MCP
Search or read in Gemini Notebook → copy evidence → switch to Claude → paste context → return to the notebook when evidence is missing.
→
With MCP
Claude calls Gemini Notebook for targeted evidence, receives the relevant result, then reasons, writes or routes the next step without a manual copy/paste handoff.
Use MCP when…
your source library is too large to paste repeatedly;
you need repeated source-grounded queries during one workflow;
cross-notebook comparison or artifact creation is part of the task;
Claude Code needs research context while it works.
Skip MCP when…
one short document can be attached once;
you only need a single Notebook answer;
the notebook contains sensitive data not approved for this architecture;
the setup and authentication overhead exceeds the handoff you are trying to remove.
Retrieval quality
Prepare the notebook before you automate it
MCP does not repair a noisy source collection. Better notebook structure usually produces more targeted retrieval and smaller, easier-to-verify handoffs.
Keep one project or decision domain per notebook
Use descriptive source names instead of generic filenames
Remove obsolete or duplicate versions before synthesis
Separate primary evidence from background material
Ask narrow retrieval questions before broad synthesis
Verify source IDs or citations before Claude interprets
Capabilities depend on the community server version, your Gemini Notebook plan, the client you use, and which tools you allow Claude to call.
NotebookLM MCP capabilities and limitations
Capability
Typical support
Practical use
List, create and organize notebooks
Supported
Build a project vault without opening each notebook manually.
Add URL, text, Drive and local-file sources
Conditional
Local files must exist on the machine running the server; remote file transfer needs separate infrastructure.
Query one notebook with source grounding
Supported
Retrieve cited evidence before Claude interprets or reformats it.
Compare themes across notebooks
Supported in featured package
Build contradiction maps, evidence matrices and cross-project synthesis.
Start Web or Drive research
Account dependent
Launch research and import selected sources when the feature is available to the account.
Create Studio outputs
Feature dependent
Trigger audio, video, reports, flashcards, infographics, mind maps or slide decks supported by the package and account.
Run batch operations and pipelines
Server dependent
Chain research, querying and artifact creation—but not unattended scheduling by itself.
Guarantee zero hallucinations
Not possible
Retrieval and citations improve verification; they do not eliminate model error.
Provide an enterprise-safe multi-user connector
Not turnkey
Requires authentication, HTTPS, isolation, governance and independent security review.
Context-window warning: the featured package currently documents 43 MCP tools. In current Claude Code, Tool Search is enabled by default, so full tool definitions are deferred until Claude needs them; only tool names and server instructions load at session start. That substantially reduces schema overhead, but retrieved tool results and long conversations still consume context. Use /context to inspect usage and /mcp to disable servers you are not using.
MCP token economics
Does NotebookLM MCP save Claude tokens?
Sometimes. The right comparison is not “MCP versus zero overhead.” It is targeted retrieval versus the source material you would otherwise place in Claude’s context.
The practical rule
MCP creates a context advantage when avoided bulk source context is larger than the tool overhead + retrieved evidence + conversation history required to complete the task.
Where MCP can save context
A long report library stays in Gemini Notebook while Claude receives only evidence relevant to the current question. This is especially useful when the alternative is repeatedly attaching or pasting large documents.
Where MCP becomes a context tax
Broad queries can return oversized results; long sessions keep accumulating history; and unnecessary servers still add names, instructions or used tool schemas to context.
Token efficiency patterns for NotebookLM MCP
Workflow pattern
Context behavior
Efficiency
Paste a very large document into Claude for every task
Bulk source material repeatedly enters Claude context
Usually low
Keep the library in Gemini Notebook and retrieve only relevant evidence
Claude receives a smaller evidence packet for the current question
Often better
Claude Code with default Tool Search
MCP tool definitions are deferred and loaded on demand Anthropic docs ↗
Long multi-task session without clearing stale history
Old conversation content is repeatedly processed
Usually low
Measure your own workflow instead of trusting a generic token estimate
Run /context before and after the workflow to see what is occupying Claude Code’s context.
Run /mcp to review configured servers and disable ones you are not actively using.
Run /usage when you need current token-usage visibility, then compare a direct-document test with an MCP-retrieval test.
Watch oversized tool results. Claude Code currently warns when an MCP tool output exceeds 10,000 tokens and applies a 25,000-token default maximum unless configured otherwise. Anthropic docs ↗
Subscription usage is not the same as API billing. Claude subscription users mainly experience usage windows and context limits; API and cloud-provider users can also incur direct per-token charges. For that reason, this guide does not hard-code a “cost per MCP session.” Measure your workflow and check current Anthropic pricing if you are billed per token.
Claude Code documentation currently states that Tool Search is on by default and MCP tool definitions are deferred; exceptions can occur with unsupported models, some third-party hosts, or configurations that explicitly disable Tool Search. Source: Anthropic ↗
High-value workflows
Four use cases where the bridge earns its complexity
These are the workflows where source grounding, repeated retrieval and Claude’s downstream reasoning or production are meaningfully complementary.
1 · Research synthesis
Keep a literature or evidence library in Gemini Notebook, retrieve claims and citations on demand, then ask Claude to compare findings, expose uncertainty and draft a decision-ready synthesis.
2 · Documentation / knowledge-base RAG
Use Gemini Notebook as the searchable source vault while Claude Code retrieves the specific documentation needed for implementation, debugging or design decisions.
3 · Cross-notebook analysis
Compare evidence across project notebooks, build contradiction maps, and surface where sources agree, disagree or leave unanswered questions.
4 · Research → deliverable
Retrieve verified evidence first, then hand it to Claude for a report, slide brief, content plan or other deliverable without repeatedly copying source material between tools.
Operating principle
Let Gemini Notebook own the evidence; let Claude own the reasoning and execution.
MCP is most useful when it removes the handoff between those two roles while keeping the evidence trail visible.
Trust layer
Security and privacy: what the setup actually exposes
MCP is a protocol—not a security certification. The safety of this workflow depends on the server code, its authentication, its host, and the permissions you grant.
Do not publish the local endpoint
The package’s HTTP mode does not become secure merely because it is reachable. Never expose it to the public internet without HTTPS, strong caller authentication and a reviewed gateway.
Browser-session credentials matter
The server authenticates with a Google browser session. Anyone who controls an unprotected server may be able to operate the active Notebook account.
Use a dedicated test account
For experimentation, isolate the workflow from your primary personal or client account. Keep only non-sensitive test notebooks in that environment.
Start read-only in practice
Your first prompts should list and query notebooks without creating, deleting, sharing or downloading anything. Add write actions only after verifying behavior.
Treat retrieved content as untrusted
Notebook sources can contain malicious or irrelevant instructions. Anthropic warns that MCP servers fetching external content can expose users to prompt-injection risk. Treat retrieved text as evidence—not as authority to change tool permissions or system instructions. Anthropic guidance ↗
Confirm write actions separately
Keep read/query steps distinct from create, delete, share and download actions. When retrieved content influences the next tool call, require explicit user confirmation before an irreversible or externally visible action. MCP security guidance ↗
Review every update
Community projects can add tools or change authentication. Read the changelog before upgrading and rerun nlm doctor afterward.
Keep regulated data out by default
Do not use this experimental bridge for health, legal, financial-client or confidential corporate information unless your organization has approved the complete architecture.
Diagnostic guide
NotebookLM MCP troubleshooting
Start with the recovery ladder, then open the symptom that matches what you see.
Authentication recovery ladder
Run nlm login --check.
If expired or inconsistent, run nlm login again.
If browser handoff fails, fully quit Chrome / Edge / Brave and retry so the dedicated debug profile can bind correctly.
Run nlm setup list to confirm the client registration.
Run nlm doctor for package, auth and configuration diagnostics.
On WSL2, verify you are using a current release with Windows-browser authentication support before attempting custom networking workarounds.
uv, nlm or notebooklm-mcp is not recognized
Close and reopen the terminal so the PATH update loads, then confirm installation with uv tool list. On Windows, also confirm the install location is on PATH and reopen PowerShell after installation. Claude Desktop extension users can reinstall the current .mcpb file from Settings → Extensions without using the CLI.
The browser login loops, cannot bind, or expires
Run nlm login --check, then rerun nlm login. If the connector reports that it cannot connect to the browser, fully quit the running Chromium-family browser and retry; recent releases include fixes for Chrome/Edge handoff and orphaned-login deadlocks. Complete any Google security prompt interactively. Google can require reauthentication at any time.
Claude does not show the NotebookLM tools
Claude Desktop: open Settings → Extensions, confirm the extension is enabled, review its logs, and fully restart the app. Claude Code: run nlm setup list, nlm doctor, and use /mcp to confirm the server loaded.
Claude shows duplicate NotebookLM tools
You probably have two older NotebookLM MCP packages or two client configurations active. Remove the legacy server and keep only one current configuration. Duplicate tool names can cause Claude to call the wrong server.
Claude chooses the wrong tool or becomes less focused
The featured package exposes many tools. Disable NotebookLM when it is not needed, or restrict the enabled tools to the current task. Start prompts with a clear instruction such as “Use only read/query tools; do not create, delete or share.”
A batch job stalls or NotebookLM returns a limit error
Reduce the batch size, split the workflow into stages, and confirm the same operation works in the Gemini Notebook interface. MCP does not bypass plan quotas or service-side limits. See the site’s NotebookLM troubleshooting guide for quota-specific fixes.
A remote custom connector cannot reach the server
Anthropic’s cloud cannot reach localhost or a private address on your computer. Use the local Desktop/Claude Code mechanism, or build a public HTTPS endpoint with authentication. A tunnel alone is not sufficient.
Remote file uploads or downloads point to the wrong computer
File paths refer to the machine hosting the MCP server—not the browser or phone running Claude. Use local operation for file-heavy workflows, or build a separate secure file-transfer layer.
Three useful starting prompts
Use the bridge for evidence—not merely convenience
These prompts begin with read-only actions and force a visible separation between retrieved evidence and Claude’s interpretation.
01 · Evidence-first answer
Use the NotebookLM MCP tools to query the notebook “[NOTEBOOK NAME]” for evidence relevant to this question: [QUESTION].
Return:
1. The strongest source-backed findings with citations or source identifiers
2. Important qualifications or missing evidence
3. Contradictions among the sources
4. A separate section labeled “Claude’s interpretation”
Do not create, edit, delete, share, or download anything.
02 · Cross-notebook contradiction map
Compare these notebooks using the NotebookLM MCP tools: [NOTEBOOK A], [NOTEBOOK B], and [NOTEBOOK C].
Build a table with:
• Claim or decision question
• Evidence supporting it
• Evidence challenging it
• Which notebook and source each item came from
• Confidence level based only on the retrieved evidence
• What additional source would resolve the disagreement
Query only. Do not modify the notebooks.
03 · Research-to-deck handoff
Query “[NOTEBOOK NAME]” and create a source-grounded presentation brief for [AUDIENCE].
Include:
• One-sentence decision or thesis
• 7-slide narrative arc
• Evidence required on each slide
• Exact source references for every factual slide
• Counterargument slide
• Final recommendation and unresolved risks
Do not generate the deck yet. First return the evidence map and ask me to approve the narrative.
Multi-AI Systems
Move from a working MCP bridge to a reusable orchestration system
The full Multi-AI package turns source retrieval into repeatable research, handoff, quality-control and production workflows across Gemini Notebook, Claude, ChatGPT and Gemini.
NotebookLM MCP is a community-built Model Context Protocol server that lets MCP clients such as Claude Desktop and Claude Code call selected NotebookLM functions. Depending on the implementation, Claude may list notebooks, query sources, add sources, create Studio outputs, or download artifacts.
How do I install NotebookLM MCP in Claude Desktop?
Download the current .mcpb release, open Claude Desktop Settings → Extensions → Advanced settings → Install Extension, select the file, restart Claude Desktop, and ask Claude to list your notebooks.
Can I use nlm setup add claude-desktop?
Yes. Current releases support it again. The .mcpb extension remains the simplest Claude Desktop path, while nlm setup add claude-desktop is useful for CLI-managed, profile-aware configuration. Fully quit the target Claude Desktop profile before the CLI writes its configuration.
Why does Claude Desktop not see NotebookLM MCP?
Open Settings → Extensions, confirm the extension is enabled, review its logs, fully quit and reopen Claude Desktop, and reinstall the latest .mcpb file if the extension failed to load.
Can I connect NotebookLM to Claude without coding?
Yes. The one-click desktop extension does not require programming. The CLI route uses copy-and-paste commands rather than code and is better when you need upgrades, diagnostics, or migration from an older package.
Is NotebookLM MCP free?
The community packages covered here are open source. You still need access to Claude and NotebookLM, and each service applies its own plan limits, usage quotas, and feature availability.
Does NotebookLM MCP save Claude tokens?
It can when targeted retrieval replaces large source documents that would otherwise enter Claude context. It is not automatically cheaper: retrieved results, conversation history and used tool schemas still consume context. In Claude Code, run /context to inspect context use and /mcp to disable servers you do not need.
Is this an official NotebookLM connector?
No. Google offers official Gemini Notebook Enterprise APIs, but the consumer integrations covered here are community-built and rely on undocumented interfaces and browser-session authentication.
Can I use Claude on the web instead of Claude Desktop?
Claude supports remote custom connectors across its clients. However, the featured NotebookLM package is designed primarily for local, single-user operation and does not provide a secure, turnkey remote service. Use Claude Desktop or Claude Code locally unless you can build and review a proper hosted architecture.
Do I need a paid Claude plan?
Desktop extensions are currently available to Claude Desktop users. Claude Code requires a supported Claude subscription or API billing. Remote custom connectors are available across Claude plans, with plan-specific limits; Free users are limited to one custom connector.
Will the connection stop Claude from hallucinating?
No. It can retrieve more relevant, source-grounded evidence and make verification easier, but either model can still omit context, misread evidence or make an unsupported inference. Require citations and keep retrieval separate from interpretation.
Can the bridge create Audio Overviews, slide decks and videos?
The featured package exposes Studio-creation tools, but actual availability depends on the package version, your Gemini Notebook account, language, plan, age requirements and current product rollout. MCP cannot unlock features your account does not have.
Can it run research automatically while I sleep?
Not by itself. MCP exposes tools to Claude. Unattended scheduling requires an additional agent, scheduler, always-on host, failure handling and security controls. Keep the article’s local setup interactive unless you deliberately build that extra layer.
Why does this guide use both Gemini Notebook and NotebookLM?
Google renamed NotebookLM to Gemini Notebook in July 2026. Existing notebooks and the underlying product continue, while community repositories and user searches still commonly use the NotebookLM name.
Last checked: August 9, 2026. GitHub marked v0.9.7 (released August 6, 2026) as the latest release at the time of this check. The repository is now gemini-notebook-mcp-cli; the installable package remains notebooklm-mcp-cli. The connector uses undocumented Gemini Notebook interfaces and can change without notice, so confirm the latest release and changelog before installation.