Notte SOP

Appearance

Usage ​

Minimal example: Python SDK agent (cloud sessions) ​

The simplest Notte operation is spinning up a cloud browser session and running an AI agent against it with a natural-language task description.

python
import os
from notte_sdk import NotteClient

client = NotteClient(api_key=os.getenv("NOTTE_API_KEY"))

with client.Session(open_viewer=True) as session:
    agent = client.Agent(
        session=session,
        reasoning_model="gemini/gemini-2.5-flash",
        max_steps=30,
    )
    response = agent.run(task="Extract the top 5 story titles from news.ycombinator.com")
    print(response)

open_viewer=True opens a live replay link so you can watch the session in your browser. Set it to False for headless production runs.

CLI workflow: navigate, observe, act ​

The CLI enforces a strict loop: always observe live DOM elements first, then act using the IDs returned by observe. Never guess selectors.

bash
# 1. Authenticate (one time)
notte auth login

# 2. Start a cloud browser session
notte sessions start

# 3. Navigate to a page
notte page goto "https://news.ycombinator.com"

# 4. Observe the live DOM to get valid element IDs
notte page observe

# 5. Scrape structured data using natural language
notte page scrape --instructions "Extract the top 5 story titles and their point counts"

# 6. Stop the session when done
notte sessions stop

Key rules ​

  • Always stop sessions when done. Leaving sessions running consumes credits. Use try/finally in Python code to guarantee cleanup even on errors.
  • CLI: never guess element IDs. Always run notte page observe first and act only on IDs it returns. The docs call this the required workflow loop.
  • MCP usage: the MCP server wires Notte directly into Claude or Cursor. The host LLM calls Notte tools without you writing any session management code.
  • Credentials: store secrets in Notte Secret Vaults rather than passing them as plain text in prompts. Vault credentials are injected into sessions without being forwarded to the LLM.

Benchmarks ​

Notte reports 79.0% LLM evaluation accuracy with 47-second average task completion on standard web agent benchmarks, outperforming competing open-source frameworks at time of publication.