browser-use / browser-harness

The skill (itself named browser-use) connects over a single WebSocket to a Chrome or Chromium instance you already have running with --remote-debugging-port open, and lets an agent click, type, navigate, and work through logged-in sessions or JS-heavy, bot-protected pages that a plain HTTP fetch can't reach. The harness is small by its own account — roughly 1k lines across four files — split between the install script, the day-to-day SKILL.md, a protected core package, and agent_helpers.py, a file the agent itself is meant to keep editing.

It's built for someone who wants an agent driving a real, already-configured browser — their own cookies and logins included — rather than a disposable sandboxed one. The README is explicit that this is the point: when the agent hits something the harness can't yet do, it's supposed to write the missing helper itself and file new "domain skills" for sites like GitHub, LinkedIn, or Amazon into a local config folder as it learns them. None of that ships in the repository — what you end up with after a few sessions depends on what your own agent chose to write, not on anything we can preview for you here.