Skip to main content
Glama

mcp-nlm

A small stdio proxy for notebooklm-mcp-cli. Requires Node.js 22+ and uvx on PATH. The proxy is distributed through npm; uvx runs only the upstream NotebookLM server.

npx -y --prefer-online mcp-nlm@latest
npx -y --prefer-online mcp-nlm@latest --version

MCP client configuration:

{
  "mcpServers": {
    "notebooklm": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "mcp-nlm@latest"]
    }
  }
}

To run directly from a checkout:

node cli.js

The proxy launches uvx --from notebooklm-mcp-cli@latest notebooklm-mcp on demand. It forwards MCP requests, results, notifications and server requests without mapping tool names, arguments or schemas. On restart it replays the client's initialization handshake, hiding only that internal handshake response.

Each client request selects an idle delay: initialize and tools/list select one minute; all other requests select thirty minutes. The latest request wins. The timer starts after outstanding requests finish. Notifications do not extend it. Closing the client connection closes the upstream process.

One exception to response transparency is authentication recovery. When a tool returns an explicit authentication error, the proxy calls upstream refresh_auth once (concurrent failures for the same profile share the attempt). It preserves the original result and appends a text block reporting whether refresh succeeded. It never retries the original operation, changes the active profile, or starts an interactive login. Explicit client calls to refresh_auth pass through unchanged. Refresh has a sixty-second response timeout; expiry does not cancel upstream's ongoing recovery. Upstream may launch its managed browser in the background.

Working directory: <OS temp>/uvx/nlm. NotebookLM data and browser profiles: <OS temp>/uvx/nlm/data, overridable with NOTEBOOKLM_MCP_CLI_PATH. The uv cache retains its existing configuration. Browser detection stays upstream's default.

The latest package is resolved when upstream starts, not while it is running. Authentication uses the same data directory:

uvx --from notebooklm-mcp-cli@latest nlm login --storage file

Set NOTEBOOKLM_MCP_CLI_PATH to the data directory above before running login. Set the same variable for the MCP client if you choose a different directory. Check login with nlm login --check or explicitly refresh with nlm auth refresh, using the same uvx --from notebooklm-mcp-cli@latest prefix and data directory. Data in the OS temporary directory can be removed by system cleanup.

NotebookLM access uses an unofficial upstream integration and your Google account. Account features, quotas and any subscription charges are controlled by Google. This proxy adds no API service or per-request fee.

Upstream process state (such as in-memory jobs) does not survive idle shutdown. Capabilities and tool schemas belong to upstream and can change after an update; reconnect the MCP client if an update changes its advertised capabilities.

Development: npm run check. Disposable experiments belong in the ignored tmp/ directory. Build and verify the npm archive: npm run release:prepare. GitHub Releases publish to npm through Trusted Publishing (OIDC).

License: MIT.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      C
      maintenance
      MCP server that provides ~39 tools to interact with Google NotebookLM's internal API, enabling notebook management, source addition, chat queries, and Studio artifact generation through stdio, HTTP, and SSE transports.
      39
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables local MCP clients to access tools from a remote FastMCP server over stdio, handling OAuth authentication and Streamable HTTP communication.
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Bridges stdio MCP clients to shared, warm per-project backend instances over streamable HTTP, so repeated client sessions skip cold starts and transparently survive backend restarts. Heartbeat leases, idle-TTL reaping, and per-request deadlines keep per-project instances alive and shared across multiple clients without hanging the caller.
      2
      MIT