🌏 中文版
notebooklm-py is currently the most complete unofficial Python API for Google NotebookLM. It’s not just a simple scraper or Selenium wrapper — it directly reverse-engineers Google’s internal RPC protocol, enabling programmatic access to all of NotebookLM’s features, including capabilities that the web UI doesn’t even expose.
Core Technology: Google batchexecute RPC
Under the hood, the NotebookLM web app uses Google’s internal batchexecute RPC protocol. This isn’t unique to NotebookLM — Google Photos, Google Translate, and other products all use the same mechanism.
The request structure looks like this:
POST /_/LabsTailwindUi/data/batchexecute
f.req = [[[rpc_id, json_params, null, "generic"]]]Key design elements:
- RPC ID: Each operation maps to a 6-character identifier (e.g.
wXbhsf), discovered through reverse engineering - Position-sensitive parameters: The payload isn’t key-value pairs but nested JSON arrays where parameter meaning is determined by index position. Missing positions must be filled with
None - Anti-XSSI protection: Responses are prefixed with
)]}'\n, which must be stripped before parsing - CSRF Token: Passed as the
atparameter using theSNlM0evalue found in the page source
There is no official documentation for this protocol. The developer’s approach is to open Chrome DevTools, filter for batchexecute requests, perform actions on the web page, then reverse-engineer the RPC IDs and payload structures from the intercepted requests one by one.
Authentication: Playwright Browser Cookie Extraction
Google doesn’t provide an OAuth scope or API key for NotebookLM, so notebooklm-py takes a pragmatic approach: it uses Playwright to open a browser for the user to log in, then captures the session cookies.
pip install "notebooklm-py[browser]"
playwright install chromium
notebooklm loginThe flow is: Playwright launches Chromium → the user manually logs into Google → the program captures and stores cookies locally → all subsequent API requests include these cookies to simulate an authenticated session.
This means the authentication lifetime is tied to the Google session cookie’s validity period — once it expires, you need to log in again.
Features: More Than a Web UI Mirror
Beyond replicating every web UI operation (creating notebooks, adding sources, chat Q&A, generating audio overviews, etc.), notebooklm-py also provides capabilities the Web UI can’t:
- Batch downloading of artifacts (audio, video, slides)
- Structured export: Quizzes and flashcards can be exported as JSON / Markdown / HTML, not just viewed in the interface
- Mind map hierarchical data extraction: Retrieve structured tree data
- Editable PPTX downloads: The web UI only gives you PDFs; here you get PowerPoint files
- Programmatic permission management: Batch-configure sharing and access permissions
All content generation goes through a single CREATE_ARTIFACT RPC method, using different type codes to distinguish output types (audio, video, slides, quizzes, etc.). Long-running tasks return a task ID, with completion status tracked via polling.
Three Ways to Use It
CLI is ideal for quick operations and scripting:
notebooklm create "Research Project"
notebooklm source add "https://example.com"
notebooklm generate audio "make it engaging" --wait
notebooklm download audio ./podcast.mp3Python API is async-based, suitable for application integration:
async with await NotebookLMClient.from_storage() as client:
nb = await client.notebooks.create("Research")
await client.sources.add_url(nb.id, "https://example.com", wait=True)
result = await client.chat.ask(nb.id, "Summarize this")AI Agent skill can be installed into Claude Code, Codex, or OpenClaw:
notebooklm skill installOverall Architecture
User (Python / CLI / Agent)
│
NotebookLMClient (async session)
│
RPC Method Layer (RPCMethod enum → 6-char RPC ID)
│
HTTP POST → /_/LabsTailwindUi/data/batchexecute
(cookies + CSRF token + nested JSON payload)
│
Google NotebookLM Backend
│
Response Parsing (strip anti-XSSI → parse chunked JSON → dataclass)Overall Assessment
notebooklm-py’s core tradeoff is clear: it trades reverse engineering for full programmatic access, at the cost of stability depending on Google not changing its internal protocol.
Good fit for: prototyping, research, batch processing (e.g. generating podcast audio for 20 documents at once), or integrating NotebookLM into AI agent workflows.
Not a good fit for: production environments or services with strict stability requirements. If Google changes a single RPC ID, the entire feature breaks.
Currently at 9.1k stars and 650+ commits, it’s the most mature option in this space. If you need programmatic access to NotebookLM, this is essentially your only path.
References
Loading...