Notebooklm Rpc

A pure-TypeScript Model Context Protocol server for Google NotebookLM. Speaks Google's undocumented batchexecute RPC protocol directly — no Python runtime, no notebooklm CLI, no shelling out. Deploys cleanly to Vercel Node functions.

Companion / sibling project to notebooklm-mcp, which wraps the notebooklm-py CLI as a subprocess. Pick this one if you want zero Python dependencies and seamless serverless deployment.

What's in the box

• Native RPC client — encoder/decoder for the batchexecute wire format ported faithfully from notebooklm-py's _core.py and rpc/.
• Native chat client — POSTs to the streaming GenerateFreeFormStreamed endpoint and parses the chunked response.
• 51 MCP tools spanning notebooks, sources, chat, artifacts (all 9 generation types + downloads), notes, research, sharing, and language settings.
• Auth refresh — concurrent-safe re-scrape of SNlM0e / FdrFJe on 401/403.
• Stdio + Streamable-HTTP transports. Vercel adapter included.
• Strict TypeScript: exactOptionalPropertyTypes, noUncheckedIndexedAccess, verbatimModuleSyntax, full Zod input validation.

Quick start

npm install
npm run build

# Provide auth (one of the three options below) — see "Authentication".
export NOTEBOOKLM_STORAGE_PATH=/Users/me/.notebooklm/storage_state.json

# Run the server (stdio, default).
node dist/index.js

Hooking up Claude Desktop / Cursor / Windsurf

{
  "mcpServers": {
    "notebooklm": {
      "command": "node",
      "args": ["/absolute/path/to/notebooklm-mcp-rpc/dist/index.js"],
      "env": {
        "NOTEBOOKLM_STORAGE_PATH": "/Users/me/.notebooklm/storage_state.json"
      }
    }
  }
}

Remote / HTTP

MCP_TRANSPORT=http PORT=3000 node dist/index.js
# POST JSON-RPC to http://localhost:3000/mcp
# GET http://localhost:3000/healthz for a liveness check

Vercel

Push to Vercel — vercel.json and api/mcp.ts are already wired:

vercel deploy
# Endpoint: https://<your-project>.vercel.app/mcp

Set NOTEBOOKLM_AUTH_JSON (the inline contents of storage_state.json) as a project env var. The Vercel Node runtime supports everything we need; no custom container.

Authentication

You need a valid Google session. The simplest path is to bootstrap it once with the Python notebooklm CLI, then point this server at the resulting file:

pip install "notebooklm-py[browser]"
playwright install chromium
notebooklm login        # writes ~/.notebooklm/storage_state.json

Now choose one of:

Variable	Use when
NOTEBOOKLM_STORAGE_PATH	Local dev. Path to storage_state.json.
NOTEBOOKLM_AUTH_JSON	Serverless / CI. Inline JSON contents of the same file.
NOTEBOOKLM_COOKIES + NOTEBOOKLM_CSRF_TOKEN + NOTEBOOKLM_SESSION_ID	You already have tokens from another pipeline.

Cookies do expire. When they do, auth_status will return an isError: true result with reason: "auth" and a reminder to re-run the login step.

Tool catalog (51)

Group	Tools
Auth / status	auth_status
Notebooks	notebook_list, notebook_create, notebook_rename, notebook_delete, notebook_raw
Sources	source_list, source_add_url, source_add_text, source_delete, source_rename, source_refresh, source_fulltext, source_guide, source_wait
Chat	chat_ask, chat_history, chat_last_conversation_id, chat_configure
Artifacts	artifact_list, artifact_get, artifact_delete, artifact_rename, artifact_wait
Generation	generate_audio, generate_video, generate_cinematic_video, generate_report, generate_quiz, generate_flashcards, generate_infographic, generate_slide_deck, generate_data_table, generate_mind_map, revise_slide
Downloads	download_artifact (one tool, dispatches by artifact type — writes to outputPath or returns inline base64/text)
Notes	note_list, note_create, note_update, note_delete
Research	research_start, research_poll, research_wait, research_import_all
Sharing	share_status, share_set_public, share_set_view_level, share_add_user, share_remove_user
Language	language_get, language_set

Architecture

src/
├── auth.ts              # storage_state → cookies → CSRF + session-id
├── rpc/
│   ├── types.ts         # method IDs + enum codes (source of truth)
│   ├── encoder.ts       # encode_rpc_request, build_request_body
│   ├── decoder.ts       # strip_anti_xssi, parse_chunked_response, extract_rpc_result
│   └── client.ts        # RpcClient — fetch + auth refresh + chat streaming
├── api/
│   ├── _helpers.ts      # asString / getPath / sourceIdsTriple etc.
│   ├── notebooks.ts     # list/create/rename/delete/raw
│   ├── sources.ts       # add_url/add_text/list/delete/rename/wait/fulltext/guide
│   ├── chat.ts          # ask (streaming), history, configure
│   ├── artifacts.ts     # list/wait + 11 generation paths + download URL extraction
│   ├── notes.ts         # CRUD over GET_NOTES_AND_MIND_MAPS
│   ├── research.ts      # start/poll/import/wait
│   ├── sharing.ts       # public link, view level, per-user
│   ├── settings.ts      # language get/set
│   └── client.ts        # NotebookLMClient — namespaced facade
├── tools/               # MCP tool registrations (one file per domain)
├── transport/{stdio,http}.ts
├── server.ts            # createServer, registerAllTools
├── config.ts            # env-driven config (zod-validated)
├── errors.ts            # AuthError, RateLimitError, ProtocolError, …
├── logger.ts            # stderr-only structured logger
└── index.ts             # bin entry
api/mcp.ts               # Vercel function adapter

Limitations (v0.1)

• No file uploads. Adding PDFs / audio / images via the resumable upload protocol is not implemented yet — use source_add_url (web URLs and YouTube) or source_add_text. File support tracked under _core.py's o4cbdc method.
• Sharing payload parsing is shallow — share_status returns the raw RPC payload alongside best-effort fields. Public-link toggle and per-user invite/remove all work end-to-end.
• Slide deck downloads require EXPORT_ARTIFACT (handled automatically by download_artifact); other artifact types extract URLs from the artifact_list response.

Why two MCPs?

	notebooklm-mcp	notebooklm-mcp-rpc
Python runtime needed	Yes (CLI subprocess)	No
Vercel-deployable	Only with custom container	Plain Node function
Tracks upstream protocol fixes	Automatic	Manual (rare; method IDs change a few times a year)
Setup steps	install Python + CLI + login	login once, then JS-only

This project (-rpc) is the right pick when you want an autonomous, single-runtime deploy.

License

MIT — see LICENSE.