← MCP server and skill grades

notebooklm-mcp

PleasePrompto/notebooklm-mcp

C
Type
MCP server
Scanned
Nov 12, 2026
Version
00179986ff6d
Files read
49

This MCP server automates a real Chrome browser logged into your Google account so AI coding agents can chat with Google NotebookLM notebooks, add web or text sources, and generate or download audio overviews. Its biggest risk is that tool calls accept any http or https address as a notebook URL, allowing a tricked AI agent to steer your browser to an attacker's site to leak saved session data, reach internal network services, or save malicious files anywhere on your computer.

What it can reach

Browser
Launches and controls Chrome or bundled Chromium with a persistent profile logged into your Google account to open pages, type questions, add sources, read citations, and download audio.
Secrets
Saves unencrypted Google account cookies and session storage to JSON files and a Chrome profile in your user data folder, and optionally reads LOGIN_EMAIL and LOGIN_PASSWORD from environment variables.
Files
Reads and writes its notebook library, settings, and browser profiles in your user data and config folders, saves downloaded audio files to any requested directory, and deletes matching caches, logs, Claude project history, and Trash items via cleanup_data.
Network
Connects to Google Accounts, NotebookLM, and any URL supplied as a notebook_url through Chrome, and optionally runs a local HTTP server on port 3000 when HTTP transport is enabled.

Findings

1 high3 medium1 low
  1. HighUnvalidated notebook_url allows arbitrary browser navigation, session data leak, SSRF, and arbitrary file writes

    src/session/session-manager.ts:76

    Tools such as ask_question, get_audio_status (which is marked read-only and often auto-approved by AI clients), add_source, and download_audio accept a notebook_url parameter (or use a URL saved via add_notebook or update_notebook) that is only checked to start with 'http'. If prompt injection in a webpage, issue, or document tricks the AI agent into passing an attacker's URL or an internal address, Chrome opens that URL using your persistent profile. In addition, BrowserSession.restoreSessionStorage() in src/session/browser-session.ts matches the page origin against notebook_url rather than https://notebooklm.google.com, injecting your saved Google/NotebookLM sessionStorage directly into the attacker's page, and download_audio saves whatever file and filename the attacker's page triggers into any destination_dir on your computer (such as ~/.ssh or editor config folders), enabling code execution.

    Fix: Parse all notebook URLs (both per-call notebook_url and library URLs in add_notebook/update_notebook) with new URL() and require protocol 'https:', hostname 'notebooklm.google.com', and a '/notebook/' path prefix; hardcode 'https://notebooklm.google.com' in restoreSessionStorage(); and restrict download_audio file extensions and overwrites.

  2. MediumHTTP transport has no authentication and no Origin or Host check against DNS rebinding

    src/transport/http.ts:140

    When the server is started with --transport http (or NOTEBOOKLM_TRANSPORT=http), the /mcp endpoint requires no password or token, checks neither Origin nor Host headers (enableDnsRebindingProtection is not enabled on StreamableHTTPServerTransport), and places no size limit on request bodies. Even on the default 127.0.0.1 address, any malicious website you visit in a browser can use DNS rebinding to open a session and call tools to read your private NotebookLM notebooks, wipe data, or download files, and binding to 0.0.0.0 (as shown in the documentation examples) exposes full unauthenticated access to anyone on the network.

    Fix: Enable enableDnsRebindingProtection with strict allowedHosts and allowedOrigins on StreamableHTTPServerTransport, require a Bearer authentication token (especially when binding to non-loopback addresses), and cap request body size in readJsonBody().

  3. Mediumcleanup_data permanently deletes matching personal files in OS Trash and external app folders without enforcing preview

    src/utils/cleanup-manager.ts:250

    The cleanup_data tool always runs in 'deep' mode and ignores the 'optional' flag on categories when deleting. If an agent calls cleanup_data with confirm: true (no prior preview call is enforced by the server), it permanently deletes any file or folder in your macOS or Linux Trash whose name contains 'notebooklm'—including personal PDFs, notes, or exported audio unrelated to this server—along with Claude Code project history matching '*notebooklm-mcp*', VSCode/Cursor logs, and npx cache directories.

    Fix: Limit cleanup_data to the server's own data, config, cache, and log directories, remove system Trash and editor/Claude project scanning (or gate optional categories behind explicit flags), and require a server-issued preview token before confirm: true is accepted.

  4. MediumPlaintext Google session cookies are exported without restrictive file permissions

    src/auth/auth-manager.ts:68

    After login, AuthManager.saveBrowserState() writes all Google authentication cookies (which grant access to your full Google account, not just NotebookLM) and sessionStorage into plaintext browser_state/state.json and session.json files using default OS permissions. On shared Linux or macOS machines with a standard 0022 umask and a traversable home directory, these files are created world-readable (0644), allowing another local user on the machine to copy them and hijack your Google session.

    Fix: Create data and browser_state directories with mode 0o700 and write or chmod state.json, session.json, and login failure screenshots with mode 0o600.

  5. LowTool profiles and disabled-tools settings are not enforced when tools are called

    src/index.ts:186

    Users can configure NOTEBOOKLM_PROFILE (such as 'minimal' for read-only tools) or disabled-tools (such as 'cleanup_data,re_auth') to block destructive operations. However, SettingsManager.filterTools() only hides those tools from the tools/list response; the CallToolRequestSchema handler still executes any of the 20 built-in tools if called directly by an HTTP client or an agent host that does not block unlisted tools client-side.

    Fix: Check at the top of the CallToolRequestSchema handler that the requested tool name exists in this.toolDefinitions and return an error if it has been filtered out by the active profile or disabled-tools list.