notte

Notte Browser is the foundational browser component of the Notte agentic ecosystem. It provides a seamless interface between LLM agents and web browsing capabilities, enabling AI agents to interact with web content in a structured and efficient manner.

Key capabilities include:

• Browser Management: Managing browser sessions and windows (powered by Playwright).
• DOM Processing: Transforming web pages into structured, agent-friendly formats.
• Error Handling: Robust management of browser-related error scenarios.
• Configurable Options: Support for headless mode, user agents, proxies, and other settings.
• Security Controls: Built-in web security controls and customizable Chrome arguments.

The notte-llm package provides the following features for Notte integrations:

• LLM engine and service implementations
• Prompt management and templating
• LLM usage tracing and logging
• Structured output support

Notte can be connected to various AI agents, browser automation stacks, and developer tools. Available integrations include:

• MCP Server: Provides AI agents access to Notte via the Model Context Protocol (MCP).
• Claude Code: Integration for using Notte within Claude Code environments.
• OpenAI CUA: Connects Notte with OpenAI computer-use agents.
• Stagehand: Allows you to use Notte cloud browsers within the Stagehand framework.

Notte agents include several built-in advanced capabilities:

• Structured Output: Get type-safe responses using Pydantic models.
• Vaults & Personas: Use credentials and identities in automations.
• Visual Understanding: Analyze images and visual page elements.
• Replay & Debugging: Debug with MP4 replays of agent execution.
• Agent Fallback: Automatic recovery from script failures.

Notte offers several products to accelerate the development lifecycle:

• Agent Builder: Allows users to create automation scripts through AI-driven conversation.
• Demonstrate Mode: Enables recording of workflows to automatically generate automation scripts.
• Studio: A live, browser-enabled IDE for editing, debugging, and one-click deployment of automations.

Notte BUA (Browser-Using Agent) is a model designed to interact exclusively with web browsers. Unlike traditional Computer-Using Agents (CUA) that rely primarily on vision and reasoning to control a general computer interface, BUA leverages the DOM (Document Object Model) of a web page alongside screenshots. This combination allows for improved understanding, reasoning, and performance when performing web-based tasks.

BUA is accessed via the bua/completions endpoint.

NottePersona provides automated identity management for AI agents. It enables agents to perform web automation tasks that require unique digital identities, such as account creation and two-factor authentication (2FA), without manual intervention.

Key features of a NottePersona include:

• Unique Email Address: A dedicated mailbox with full email management capabilities.
• Phone Number: An SMS-capable number for verification and 2FA flows.
• Credential Vault: Optional secure storage for passwords and authentication tokens.
• Automated Communication: Built-in capabilities to read incoming emails and SMS messages.
• 2FA Support: Seamless handling of authentication flows via email or SMS.

A Notte session transitions through several distinct states during its lifecycle. Monitoring these states is essential for managing automation workflows and resource cleanup.

• active: The session is running and ready to accept operations (e.g., execute, observe, scrape).
• closed: The session has been terminated normally.
• error: The session encountered an error and terminated unexpectedly.
• timed_out: The session exceeded its configured timeout_minutes duration.

Notte sessions include built-in stealth features designed to help automations avoid detection by anti-bot systems. These features are active by default in every session and include:

• Clean browser fingerprints: Realistic browser signatures.
• Navigator properties: Proper webdriver, plugins, and language settings.
• Canvas fingerprinting protection: Randomized canvas signatures.
• WebGL fingerprinting protection: Unique WebGL parameters.
• Timezone and locale matching: Consistent geolocation data.
• Chrome DevTools Protocol hiding: Prevention of CDP detection.

The Vault acts as a secure intermediary between an LLM agent and a browser session. To prevent sensitive data from being exposed to LLMs, the system uses a 'dummy credential' pattern.

When an agent's intent involves a sensitive action (like FillActions requiring passwords, credit card numbers, or MFA secrets), the Vault intercepts the execution. It manually replaces the dummy credentials provided by the LLM with the actual, encrypted credentials stored in the Vault. This ensures the LLM never sees the real secrets, even though the agent can successfully complete the authenticated task.