Browser Harness connects an LLM-driven coding agent to your own Chrome browser through a single Chrome DevTools Protocol (CDP) websocket. The layer is deliberately thin: when the agent needs an action the harness does not yet have, such as uploading a file, it writes the missing helper into its local workspace mid-task, so the set of helpers grows with use. The harness core stays protected while the agent edits only its own workspace helpers.
Setup is driven by a prompt pasted into Claude Code or Codex. The agent installs the tool, registers a skill that teaches it the browser workflow, and connects to the browser once remote debugging is enabled. Local browser recordings are optional and off by default. Because it drives your real browser, it suits logged-in, personal tasks, for example downloading your latest videos from an X profile.
An MCP server exposes the same browser control helpers as MCP tools, so MCP clients such as Claude Code, Devin, and Cursor can drive the browser without a second CDP layer. For many browsers in parallel, with live previews, proxies, stealth, and CAPTCHA solving, the harness can run on Browser Use Cloud instead of local Chrome.
Features
- Self-healing helpers: the agent writes missing helpers into its workspace as it works, and they are reused on later tasks
- Thin CDP layer: one editable websocket connection to a real Chrome browser
- Protected core: the harness source stays untouched while agent-written helpers live in a separate workspace
- Agent-driven setup: a setup prompt for Claude Code or Codex installs the tool, registers the skill, and connects the browser
- Browser skill: a bundled skill teaches the agent the browser workflow
- MCP server: browser control helpers exposed as MCP tools for Claude Code, Devin, Cursor, and other clients
- Optional recordings: local browser recordings can be enabled, off by default
- Cloud scaling: run many parallel browsers with proxies, stealth, and CAPTCHA solving on Browser Use Cloud
