Back to directory

Browser Harness

A harness that connects a language model directly to a real browser over an editable Chrome DevTools Protocol connection, letting the agent write and reuse its own helper functions for browser tasks as it works.

Browser Harness

Browser Harness connects a language model to a real browser through a single, editable CDP WebSocket connection. When the agent needs a capability it does not have, it writes a missing helper function into its own local workspace, so the harness accumulates reusable helpers and improves with every task.

Setup connects an agent, such as one running in Claude Code or Codex, to a browser through Chrome's remote debugging interface; an install guide connects the agent to the browser and a SKILL.md file teaches it the browser workflow, while the harness's own source stays fixed. A companion MCP server exposes the same browser control helpers as MCP tools over stdio, so other MCP clients can drive the browser without a separate CDP integration.

For running many browsers in parallel, with live previews, proxies, stealth and CAPTCHA solving, the project can scale through a hosted cloud service. It is written in Python.

More in Browser automation

See all 88
  • Loads an unpacked Chrome extension that lets a local AI coding agent click, type and read the tabs, cookies and logins of a Chrome window you already have open.

  • Hronaut

    hronaut.dev

    A visible, persistent Electron browser that exposes durable, named agent workspaces over MCP, so a coding agent's browser session stays open and inspectable independently of any one AI conversation.

  • Prompt OS

    github.com

    A desktop AI agent for Windows that executes terminal commands, manages files, writes and runs code, browses the web and reads the screen, acting on natural-language requests rather than only answering them.