Browser Harness

Browser Harness is a lightweight, self-healing browser automation harness (~592 lines of Python) that connects LLMs directly to Chrome via a single WebSocket using the Chrome DevTools Protocol (CDP). It gives AI agents complete freedom to complete any browser task without frameworks, recipes, or rails. When a helper function is missing mid-task, the agent writes it itself — making the harness genuinely self-improving over time. It supports both local browsers and free remote browsers with proxies and captcha solving.

• Self-healing architecture: When the agent encounters a missing tool or helper function, it edits helpers.py on the fly and writes the missing capability, then continues the task without interruption.
• Direct CDP connection: Built on a single WebSocket to Chrome with no intermediary framework — just run.py, helpers.py, and a daemon bridge (admin.py + daemon.py).
• Domain skills system: Agents generate and save reusable skill files under domain-skills/ for sites like GitHub, LinkedIn, and Amazon, capturing selectors and flows so they don't have to rediscover them.
• Free remote browsers: Integrates with cloud.browser-use.com for up to 3 concurrent remote browsers with proxies and captcha solving — no credit card required.
• Minimal footprint: The entire harness is ~592 lines of Python across a handful of files, making it easy to audit, fork, and extend.
• LLM-agnostic setup: Paste the setup prompt into Claude Code or Codex to have the agent install and configure the harness against your real browser automatically.
• Open source (MIT): Freely available on GitHub under the MIT License — PRs, domain skill contributions, and bug fixes are all welcome.