BrowserCode is a coding agent that controls real browsers by writing JavaScript against the Chrome DevTools Protocol (CDP). It treats browser interaction as a coding problem: the agent adapts to each site at runtime and saves reusable scripts to .bcode/agent-workspace/ for later runs.
It is a fork of OpenCode with a vendored TypeScript port of Browser Harness, and adds one browser primitive, browser_execute(code), which runs JavaScript in-process, talks to Chrome over CDP, keeps the browser session alive across calls, and returns logs, values, and screenshots to the agent. BrowserCode is not affiliated with the OpenCode team.
You can open a terminal UI with bcode, run a task headlessly with bcode run "...", or call the hosted version through the Browser Use Cloud API. The agent can attach to your current tab, create its own local browser profile, or open a Browser Use Cloud browser and return a link to watch it.
BrowserCode works with any model reachable by API key plus every provider OpenCode supports. Current BU Bench evaluations list claude-opus-4-8 for top performance, kimi-k3 as the top open-weight model, and gpt-5.6-luna at xhigh reasoning effort for value. Anonymous usage traces are sent unless you set DONOTTRACK=1.
Features
- browser_execute: a single primitive that runs JavaScript against Chrome over CDP and persists the session between calls
- Script reuse: the agent writes browser scripts to a workspace directory for later tasks
- Three browser modes: attach to your live tab, work in a fresh local profile, or drive a Browser Use Cloud browser with a live view link
- Cloud browsers: Browser Use Cloud provides browsers with stealth, CAPTCHA solving, and proxies, limited to 3 concurrent sessions on the free tier
- Model flexibility: any API-key model plus OpenCode's provider list, configured via /connect or environment variables
- TUI and headless modes: interactive terminal UI or one-shot bcode run commands
- Hosted API: POST /api/v4/runs on Browser Use Cloud runs the agent without local setup
