browser-control

Why we included this project

Wiring an LLM agent to act on the open web usually means babysitting a Playwright or Selenium harness. This small Rust binary sidesteps that: each action is a shell subcommand that talks to Chrome over the DevTools Protocol, so you can launch a browser, snapshot the page into stable @e1/@e2 handles, click, fill, press keys, and run arbitrary JavaScript. The handles are chosen for model output, so an agent can act on `click @e3` instead of brittle selectors. When the high-level surface runs out, raw eval and cdp commands cover the rest, and a hidden daemon keeps rolling network, console, and event buffers plus per-failure traces so the agent can diagnose its own mistakes. The same commands work against a local Chrome or remote sessions from Browser Use, Steel, Hyperbrowser, or Browserbase, and because there is no LLM or MCP dependency, it drops into an existing stack as a plain shell utility.