browser-use/browser-harness

Browser Harness | Self-healing harness that enables LLMs to complete any task.

What it solves

Standard browser automation often lacks the flexibility required for complex, unpredictable web tasks. This project provides a way to connect an LLM directly to a real Chrome/Chromium browser via a websocket, giving an agent the freedom to perform tasks without the limitations of rigid automation scripts.

How it works

The harness connects to a running browser via the Chrome DevTools Protocol (CDP). It uses a unique approach where the agent can write its own missing helper functions or domain-specific skills during execution. When an agent encounters a task it cannot complete with existing tools, it writes the necessary Python code to solve it, effectively improving the harness's capabilities with every run.

Who it’s for

  • AI Agent Developers looking for a way to give their models complete freedom within a real browser environment.
  • Automation Engineers who need agents to handle complex, site-specific workflows like LinkedIn outreach or Amazon ordering.
  • Developers wanting to build self-improving web agents that learn selectors and flows through experience.

Highlights

  • Self-improving capabilities: The agent writes its own helper functions and domain skills as it encounters new tasks.
  • Direct CDP connection: Uses a single websocket to connect to Chrome, providing a thin and editable interface.
  • Domain skills: Supports reusable, site-specific skills that teach agents specific selectors and flows.
  • Cloud options: Offers a cloud service for stealth, sub-agents, or headless deployment with captcha solving included.