Notte: Reliable Browser AI agents

Notte presents itself as a full-stack framework for web automation agents. It combines scripting for deterministic steps with AI only where needed, and says this can cut costs by more than 50% while improving reliability. The project also offers a single API to develop, deploy, and scale agents and automations.

Reading notes

#

• The opensource core includes running web agents from natural language tasks, structured output with Pydantic models, and site interactions using Playwright-compatible primitives and natural language commands.
• The API service adds stealth browser sessions with CAPTCHA solving, proxies, and anti-detection.
• It also includes hybrid workflows, secrets vaults for credentials, and digital personas with unique emails, phones, and automated 2FA for account creation flows.
• The quickstart shows installation with pip install notte and patchright install --with-deps chromium.
• The local mode example uses notte.Session and notte.Agent with a reasoning model and a task prompt.
• The Python SDK example uses NotteClient, hosted browser sessions, and a drop-in change from notte objects to cli objects.
• The benchmarks table reports Notte as ranking first, with 96.6% task reliability, 47 seconds per task, and higher reported scores than Browser-Use and Convergence.
• Structured output lets the agent return data in a specified Pydantic model.
• Vaults store credentials and are used automatically when needed.
• Personas provide digital identities with unique email addresses, phone numbers, and automated 2FA handling.
• Session features include stealth options, built-in CAPTCHA solving, proxy configuration, file upload and download support, and cookie-based authentication.
• The project supports external CDP-compatible browser sessions.
• Workflows combine scripted navigation and agent reasoning in the same session.
• The scraping endpoint can extract data with structured outputs and custom instructions.
• The license is SSPL-1.0.