About this project

Browser Harness is a browser automation bridge for LLM agents. Rather than providing a fixed set of actions, it connects the agent to the user's real browser through one editable CDP websocket and teaches the browser workflow through a skill definition. During a task, if a needed helper is missing, the agent can write it in a local workspace, making the helper reusable for later work; the package's own source directory remains protected. The documented setup is intended for Claude Code or Codex using a setup prompt. The prompt asks the coding agent to install or upgrade browser-harness with uv on Python 3.12, register the output of `browser-harness skill`, connect to the browser, and ask whether local browser recordings should be enabled. Connection uses Chrome remote debugging, including the `chrome://inspect/#remote-debugging` permission step, with `install.md` as the fallback setup guide. An MCP server, `browser-harness-mcp`, exposes browser-control helpers as MCP tools over stdio. This allows MCP clients such as Claude Code, Devin, and Cursor to drive the browser without implementing a separate CDP integration. The repository shows a personal-browser workflow for logged-in tasks and links Browser Use Cloud for running many browsers in parallel with live previews, proxies, stealth measures, CAPTCHA solving, and related managed features. The example task demonstrated in the README is opening an X profile, locating the latest 20 video posts, and downloading them. Contributions mentioned include bug fixes, documentation improvements, and agent-generated domain skills.