本文へ移動
cccskills
無料GitHub で公開

notebooklm-mcp

Use when the user wants to configure, authenticate, or use Google NotebookLM via the nlm CLI or MCP.

インストール方法を見る

含まれるファイル(7)

  • SKILL.md23.5 KB
  • references/auth-guide.md7.8 KB
  • references/mcp-config.md8.1 KB
  • references/platform-quirks.md7.8 KB
  • scripts/extract_cookies_help.sh1.6 KB
  • scripts/setup_notebooklm_mcp.sh6.9 KB
  • scripts/verify_notebooklm.sh1.4 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

NotebookLM (Gemini Notebook) — CLI + MCP

Google NotebookLM has no official API. The notebooklm-mcp-cli package (nlm CLI + notebooklm-mcp server) is a third-party, unofficial client that authenticates by extracting browser cookies from a logged-in Google session and caching them. This skill covers the two headless-friendly auth methods and the MCP server wiring.

Security note: a cookies.txt file or the cached auth.json is equivalent to a Google session. The agent must never extract, read, or forward cookie values without the user's explicit consent.

PathBinaryTransportWhen to use
A. CLInlmshellAny agent with shell access. Full notebook/source/note/chat/studio management.
B. MCPnotebooklm-mcpstdio (or http/sse)MCP-native agents (Claude Code, Cursor, Devin, Gemini CLI). Exposes ~30 tools (notebook_create, source_add, chat, audio, report…).

Both paths share the same cookie cache at ~/.notebooklm-mcp-cli/profiles/<profile>/auth.json.

When to use

  • devin mcp list / claude mcp list shows notebooklm-mcp failing to list tools.

  • nlm login --check fails with ClientAuthenticationError or network_error.

  • nlm doctor reports "Browser: not found" (headless server).

  • The user wants to authenticate NotebookLM on a server without a desktop browser.

  • The user asks to configure the NotebookLM MCP server for Claude Code / Cursor / Devin / Gemini.

  • The user needs to extract Google cookies manually or authenticate via an external CDP endpoint.

  • User asks or mentions this skill in English (e.g., "use /notebooklm-mcp", "run notebooklm-mcp").

  • O usuário pede ou menciona esta skill em português (ex.: "use /notebooklm-mcp", "execute notebooklm-mcp").

When NOT to use

  • The user wants a generic "scrape Google docs" tool — NotebookLM is specifically for the NotebookLM product.
  • The user is on a desktop with Chrome installed — just run nlm login (auto mode); no skill needed.

Guardrails

  • User-performed install: the agent does not install or upgrade notebooklm-mcp-cli. If nlm/notebooklm-mcp is missing from PATH, stop and hand the user the Install block below — the user reviews it and runs it. Recommend an explicit pinned ==<VERSION>; never a bare, unversioned install.
  • Verify upstream before the user installs: confirm the package name and version on PyPI and the upstream source. Treat it as an unofficial client.
  • Human confirmation required for all auth methods: the agent must obtain explicit user consent before any cookie extraction, file import or use of a managed browser CDP endpoint.
  • No automated cookie collection: never run cookie extraction in the background or against a browser the user does not explicitly authorize.
  • Cookies are credentials: a cookies.txt file or the cached auth.json is equivalent to a Google session. Never commit, share, log, screenshot, copy to clipboard, or transmit them.
  • Delete cookie files immediately: remove cookies.txt as soon as nlm login --manual --file succeeds. Do not leave it in /tmp, the project directory, or any shared location.
  • No credential brokering: run nlm login as a black-box command. Do not read, parse, or forward the contents of cookies.txt or auth.json.
  • Prefer official auth: use desktop nlm login auto mode when a browser is available. Use manual cookie mode only on headless servers the user controls.
  • No browser data harvesting: extract cookies only from the user's own browser session; do not use extracted cookies for any purpose other than authenticating nlm.
  • Verify before trusting: run nlm login --check and nlm doctor before any notebook operation.
  • No privileged operations: do not run nlm with elevated privileges (sudo) or as root unless the user explicitly requests it and confirms the reason.
  • Isolate auth cache: prefer per-project or per-profile auth over a shared default profile when multiple users may access the environment.
  • Account risk disclosure: nlm is an unofficial client and may violate Google's Terms of Service; accounts using it can be rate-limited or suspended. Surface this risk and recommend a secondary/dedicated Google account — never push the user's primary account into it without that warning.
  • Session scope: imported cookies grant a full Google web session, far broader than NotebookLM. Treat auth.json/cookies.txt as account-level credentials, not NotebookLM-scoped tokens.
  • Data sensitivity: everything added to a notebook (sources, files, chat prompts) is sent to Google. Never add secrets, credentials, or PII as sources or prompts; confirm with the user before uploading files whose contents were not reviewed.
  • MCP transport scoping: prefer stdio (local agent only). Use --transport http/sse only when the user explicitly asks; bind to localhost and require a TLS + auth reverse proxy before any remote exposure.

Install (performed by the user, not the agent)

notebooklm-mcp-cli is a third-party PyPI distribution. The agent must not run package installs for it — if nlm/notebooklm-mcp is absent, present the block below verbatim and wait for the user to run it themselves.

Pin a known version. Replace <VERSION> with a stable release the user verified on PyPI (e.g. via pip index versions notebooklm-mcp-cli).

# Option 1: uv (recommended)
uv tool install notebooklm-mcp-cli==<VERSION>

# Option 2: pipx
pipx install notebooklm-mcp-cli==<VERSION>

# Option 3: pip (only in a virtual environment; avoid --user)
pip install notebooklm-mcp-cli==<VERSION>

Verify:

nlm --version          # 0.9.x
which nlm notebooklm-mcp

Run the diagnostic at any time:

nlm doctor             # checks install, auth, browser, AI-tool configs
nlm doctor -v          # verbose

Authentication

NotebookLM has no API key. The supported authentication method is nlm login on a machine with a browser. The agent must not extract, read, or forward Google session cookies unless the user explicitly requests a headless fallback and confirms the method.

MethodCommandRequiresBest for
Desktop auto login (preferred)nlm loginA desktop/server browser (Chrome/Chromium/Brave/Edge/Arc/Firefox)Default — preferred whenever a browser is available
OpenClaw CDPnlm login --provider openclaw --cdp-url http://127.0.0.1:18800OpenClaw-managed browser already logged in to GoogleWhen the user runs OpenClaw and confirms the CDP endpoint
Manual cookie filenlm login --manual --file cookies.txtA cookies.txt file provided by the userEmergency fallback on a headless server the user controls
Desktop auto + copy auth.jsonnlm login on desktop, then copy ~/.notebooklm-mcp-cli/profiles/<profile>/auth.jsonA desktop with a browserThe user explicitly copies the cached auth file themselves

Fallback order on headless boxes: (1) desktop auto login and copy auth.json by the user, (2) OpenClaw CDP with a managed browser the user controls, (3) manual cookie file as an emergency fallback only.

Desktop auto login

On a desktop or server with a supported browser:

nlm login              # launches a dedicated browser profile; the user logs in
nlm login --check

Prefer a specific browser:

nlm config set auth.browser chromium   # or brave, arc, edge, chrome, firefox, vivaldi, opera

This is the only auth method the agent should run autonomously.

OpenClaw CDP (requires user approval)

nlm login --provider openclaw --cdp-url http://127.0.0.1:18800

If you configured a custom OpenClaw browser profile with a different CDP port:

# Check your OpenClaw browser config
openclaw config get browser.profiles

# Use the matching CDP port
nlm login --provider openclaw --cdp-url http://127.0.0.1:<port>

The OpenClaw browser must already be logged in to Google / NotebookLM. Uses suppress_origin=True for websocket CDP commands to support managed endpoints that reject the default Origin header.

Method 1 — Manual cookie file (emergency fallback only)

The agent must not perform these steps automatically. Manual cookie extraction is an emergency fallback for a headless server the user controls. The user must create cookies.txt themselves and delete it immediately after import.

Step 1: The user extracts cookies on a machine with Chrome

  1. Open Chrome and go to https://notebooklm.google.com
  2. Make sure you are logged in to your Google account.
  3. Press F12 (or Cmd+Option+I on Mac) to open DevTools.
  4. Click the Network tab.
  5. In the filter box, type: batchexecute
  6. Click on any notebook to trigger a request.
  7. Click on a batchexecute request in the list.
  8. In the right panel, scroll to Request Headers.
  9. Find the line starting with cookie:.
  10. Right-click the cookie value and select Copy value.
  11. Paste into a text file and save as cookies.txt.

Cookie file format

The file should contain the raw cookie string from Chrome DevTools:

SID=abc123...; HSID=xyz789...; SSID=...; APISID=...; SAPISID=...; __Secure-1PSID=...; __Secure-3PSID=...; ...
  • Lines starting with # are treated as comments and ignored.
  • The file can contain the cookie string on one or multiple lines.

Step 2: Import on the server

# The user copies cookies.txt to the server, then:
nlm login --manual --file cookies.txt

# Delete the cookie file immediately after successful import
shred -u cookies.txt 2>/dev/null || rm -f cookies.txt

Step 3: Verify

nlm login --check
# ✓ Authenticated as user@example.com

nlm doctor
# Authentication: cookies present, CSRF token: yes, account: user@example.com

Tokens are cached at ~/.notebooklm-mcp-cli/profiles/default/auth.json.

Method 2 — OpenClaw CDP provider

If an OpenClaw-managed browser is already running and exposing a Chrome DevTools Protocol endpoint, nlm can read cookies from it without launching a second browser:

nlm login --provider openclaw --cdp-url http://127.0.0.1:18800
  • Uses suppress_origin=True for websocket CDP commands to support managed endpoints that reject the default Origin header.
  • The browser must already be logged in to Google / NotebookLM.
  • Increase the DevTools timeout if the endpoint is slow: nlm login --provider openclaw --cdp-url http://127.0.0.1:18800 --devtools-timeout 15

Verify the same way:

nlm login --check

Method 3 — Auto mode (desktop with browser)

On a desktop with Chrome/Chromium/Brave/Edge/Arc/Firefox installed:

nlm login              # launches a dedicated browser profile, you log in, cookies extracted

Prefer a specific browser:

nlm config set auth.browser chromium   # or brave, arc, edge, chrome, firefox, vivaldi, opera

Multi-profile auth

Multiple Google accounts are supported via named profiles:

nlm login --profile work
nlm login --profile personal
nlm login switch work
nlm login profile list
nlm login profile delete personal

The MCP server always uses the active default profile. Switching the default profile instantaneously switches the MCP server's Google account:

nlm login switch personal
nlm config set auth.default_profile personal

Auth lifecycle

ComponentDurationRefresh
Cookies~2-4 weeksAuto-refresh via headless browser (if profile saved)
CSRF tokenminutesAuto-refreshed on every request failure
Session IDsessionEmbedded in cookies

When cookies go stale:

nlm login --check        # reports stale/unverified
nlm login                # re-extract (auto mode)
nlm login --manual --file cookies.txt   # re-extract (manual mode)

Auth Security Checklist

Before any auth operation, confirm all items below. Stop and ask the user if any item is unclear.

  • The user explicitly requested this auth method and understands it involves Google session cookies.
  • The target machine is controlled by the user (personal workstation or server they administer).
  • No cookie file will be committed, logged, copied to clipboard, or left on disk after import.
  • If using --provider openclaw, the CDP endpoint is the user's own OpenClaw-managed browser.
  • If using --manual --file cookies.txt, the user provided the file and will delete it after import.
  • nlm login --check and nlm doctor pass before any notebook operation.

PATH A — CLI usage

# Notebooks
nlm notebook list
nlm notebook create --title "My Notebook"
nlm notebook get <notebook_id>
nlm notebook delete <notebook_id>

# Sources
nlm source add <notebook_id> --url https://example.com/doc
nlm source add <notebook_id> --file ./paper.pdf
nlm source list <notebook_id>

# Notes
nlm note add <notebook_id> --text "My note"
nlm note list <notebook_id>

# Chat
nlm chats start <notebook_id>
nlm chats send <notebook_id> "Summarize the sources"

# Audio overviews
nlm audio create <notebook_id>
nlm download <notebook_id> --artifact audio

# Reports
nlm report create <notebook_id> --topic "Key findings"

# Research
nlm research discover "climate adaptation strategies"

# Batch + cross-notebook
nlm batch <notebook_ids_file> --command "summarize"
nlm cross query "find mentions of X across all notebooks"

# Tags + labels
nlm tag add <notebook_id> research
nlm label add <notebook_id> source_1 priority

# Sharing + export
nlm share <notebook_id> --email collaborator@example.com
nlm export <notebook_id> --format docs

# Diagnostics
nlm doctor
nlm doctor auth-replay     # diagnose cookie replay vs browser-bound auth failures

PATH B — MCP server

B.1 Configure

The easiest way is nlm setup (covers Claude Code, Claude Desktop, Cursor, Gemini CLI, GitHub Copilot, Windsurf):

nlm setup add claude-code       # Claude Code
nlm setup add claude-desktop    # Claude Desktop
nlm setup add gemini            # Gemini CLI / Antigravity IDE
nlm setup add github-copilot    # GitHub Copilot
nlm setup add cursor            # Cursor
nlm setup add windsurf          # Windsurf
nlm setup add json              # Any other tool (interactive JSON generator)

nlm setup list                  # show supported tools + their MCP config status

For platforms not covered by nlm setup (Devin CLI/Desktop, OpenCode, Antigravity CLI, OpenClaw), use the bundled setup script or edit the config manually.

Per-platform config — critical gotchas

Each MCP client platform has its own config format. Getting field names wrong causes the server to be silently ignored (no error, just no tools). See references/platform-quirks.md for the full matrix.

PlatformConfig fileRoot keyStdio commandGotcha
Claude Code~/.claude.jsonmcpServerscommand + argstype: "stdio"
Claude Desktopclaude_desktop_config.jsonmcpServerscommand + args—
Cursor~/.cursor/mcp.jsonmcpServerscommand + argstype: "stdio" required
Devin CLI~/.config/devin/mcp_config.jsonmcpServerscommand + argsdevin mcp add CLI
Devin Desktop~/.devin/mcp_config.jsonmcpServerscommand + args—
OpenCode~/.config/opencode/opencode.jsonmcpcommand (single array)type: "local", environment not env
Antigravity IDE/CLI~/.gemini/config/mcp_config.jsonmcpServerscommand + argsClear cache on uninstall
OpenClawOpenClaw configmcp.serverscommand + argsopenclaw mcp add CLI

Top 3 silent-failure traps:

  1. OpenCode uses mcp (not mcpServers), environment (not env), command as single array (binary + args merged).
  2. OpenClaw uses mcp.servers (dotted) with transport: "stdio", managed via openclaw mcp add/set.
  3. Antigravity caches MCP servers in ~/.gemini/antigravity{,-ide,-cli}/mcp/ — must delete cache dir to uninstall.

See references/mcp-config.md for the exact JSON block per platform.

Automated setup helper

Run the bundled helper to detect all installed platforms and patch each with the correct format:

bash skills/notebooklm-mcp/scripts/setup_notebooklm_mcp.sh
# dry-run:
bash skills/notebooklm-mcp/scripts/setup_notebooklm_mcp.sh --dry-run
# target one platform:
bash skills/notebooklm-mcp/scripts/setup_notebooklm_mcp.sh --platform cursor
# use bare notebooklm-mcp binary instead of nlm wrapper:
bash skills/notebooklm-mcp/scripts/setup_notebooklm_mcp.sh --binary notebooklm-mcp
# remove:
bash skills/notebooklm-mcp/scripts/setup_notebooklm_mcp.sh --remove

Transport options

notebooklm-mcp                          # stdio (default, for local agents)
notebooklm-mcp --transport http --port 8000   # HTTP (for remote agents)
notebooklm-mcp --transport sse  --port 8000   # SSE

Env vars:

VariableDescription
NOTEBOOKLM_MCP_TRANSPORTTransport type (stdio/http/sse)
NOTEBOOKLM_MCP_PORTHTTP/SSE port
NOTEBOOKLM_MCP_DEBUGEnable verbose logging
NOTEBOOKLM_HLInterface language / locale (e.g. pt-BR, es-419)
NOTEBOOKLM_QUERY_TIMEOUTQuery timeout (seconds)
NOTEBOOKLM_BASE_URLOverride base URL for Enterprise/Workspace

Remote MCP warning: HTTP transport does not provide HTTPS, caller authentication, per-user NotebookLM accounts, or remote file transfer. Do not expose it publicly without a reverse proxy adding TLS + auth.

B.2 Verify

After configuring, restart the agent. The MCP server exposes ~30 tools. Check from the agent:

mcp_list_tools("notebooklm-mcp")
# Expect: notebook_create, notebook_list, source_add, source_list, chat_send,
#         audio_create, report_create, refresh_auth, save_auth_tokens, ...

From the shell:

bash skills/notebooklm-mcp/scripts/verify_notebooklm.sh

B.3 Auth tools exposed via MCP

ToolDescription
refresh_authReload auth tokens from the cached profile
save_auth_tokensSave cookies (fallback method)

If MCP tool calls fail with auth errors, call refresh_auth first. If that fails, re-run nlm login on the host.


Headless auth flow (decision diagram)

┌────────────────────────────────────────────────────────────────┐
│ Is an OpenClaw browser running with CDP on 127.0.0.1:18800?    │
│   YES → nlm login --provider openclaw --cdp-url http://...     │
│   NO  ↓                                                        │
│ Do you have a cookies.txt file (extracted on another machine)? │
│   YES → nlm login --manual --file cookies.txt                  │
│   NO  ↓                                                        │
│ Can you run nlm login on a desktop with Chrome?                │
│   YES → nlm login (auto mode) → copy auth.json to the server   │
│         (cp ~/.notebooklm-mcp-cli/profiles/default/auth.json   │
│            server:~/.notebooklm-mcp-cli/profiles/default/)     │
│   NO  → cannot authenticate; NotebookLM needs Google cookies   │
└────────────────────────────────────────────────────────────────┘

Troubleshooting

SymptomCauseFix
nlm login --check → ClientAuthenticationErrorCookies expired or invalidRe-extract: nlm login --manual --file cookies.txt or nlm login
nlm doctor → "Browser: not found"Headless server, no ChromeUse manual file mode or OpenClaw CDP (see Auth section)
nlm doctor → "Headless auth: not available"No saved browser profileRun nlm login once on a desktop to save the profile, or use manual mode
MCP tools fail with auth errorCached cookies staleCall refresh_auth MCP tool, or nlm login on the host
network_error on --checkCookies present but session deadRe-login; saved credentials may still be valid but session expired
Two Gemini Notebook servers configuredTool name overlap confuses agentsRemove the legacy server; keep only notebooklm-mcp
nlm setup add <tool> says "already configured"Existing entrynlm setup remove <tool> then re-add, or edit the config manually
Cookie replay fails (browser-bound auth)Google requires browser-bound sessionRun nlm doctor auth-replay to diagnose; may need auto mode with a real browser

References

  • references/mcp-config.md — Full per-platform JSON config blocks (Claude Code/Desktop, Cursor, Devin CLI/Desktop, OpenCode, Antigravity IDE/CLI, OpenClaw).
  • references/platform-quirks.md — Cross-platform MCP config quirks matrix (serverUrl vs url, mcp vs mcpServers, environment vs env, env substitution syntax, OpenClaw CDP ports).
  • references/auth-guide.md — Deep dive on cookie extraction, file format, multi-profile, OpenClaw CDP, and auth lifecycle.
  • scripts/setup_notebooklm_mcp.sh — Detects all installed platforms and patches each with the correct format (handles mcp/mcpServers, command array, environment/env, OpenClaw CLI).
  • scripts/verify_notebooklm.sh — Runs nlm doctor + nlm login --check + lists notebooks to confirm end-to-end.
  • scripts/extract_cookies_help.sh — Prints the step-by-step cookie extraction instructions for the user.
  • Authentication guide (upstream)
  • MCP guide (upstream)
  • CLI guide (upstream)
  • PyPI
  • Devin CLI MCP configuration
  • Antigravity MCP docs
  • OpenCode MCP servers
  • OpenClaw MCP tools
  • OpenClaw browser CDP
  • Cursor MCP docs

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Single owner of everything under docs/architecture/ — ADRs, architecture and design documents, and architecture diagrams. Routes each deliverable to the right engine: /mermaid-architecture for Markdown-native diagrams, /drawio-architecture for editable .drawio diagrams, and the optional archify skill for interactive standalone HTML diagrams (used only when already installed in the environment; never installed at runtime). Use whenever architecture documentation, ADRs, or architecture diagrams must be created or updated.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

Use when building a new MCP server in TypeScript, Python, or C# that exposes tools to LLMs.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

Use when reviewing code before it merges, whether written by an agent or a human.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

Use when the user asks to connect an AI agent to external apps via Composio, or when Composio CLI or MCP setup fails.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

Use when initializing or migrating an AI agent harness in a repository.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

Use when turning approved plans, specs, PRDs, or Epics into trackable GitHub Issues.

日本語の概要は準備中です。原文の説明を表示しています。

afonsoft/skills102026年10月9日 更新

afonsoft のスキルをすべて見る

このスキルの問題を報告する