A web interface for agentic browser automation. Chat-style input, visual timeline output.

A Bun-powered web app that lets you give natural language instructions to an AI agent that controls a real browser via browse. Think ChatGPT, but it actually goes and does things on the web.

Examples:

• "Research the latest pricing for Vercel, Netlify, and Cloudflare Pages and compare them"
• "Go to staging.example.com, log in, and test the checkout flow"
• "Run through the signup form and check for accessibility issues"
• "QA the landing page on mobile and desktop viewports"

┌──────────────────────────────────┐
│  Web UI                          │
│  - Chat input                    │
│  - Action timeline + screenshots │
├──────────────────────────────────┤
│  OpenCode (agent + LLM)          │
│  - Interprets user intent        │
│  - Decides which browse commands │
│  - Loops until task is complete  │
├──────────────────────────────────┤
│  browse CLI → daemon → Playwright│
│  - Navigates, clicks, fills      │
│  - Takes screenshots             │
│  - Reads page content            │
└──────────────────────────────────┘

1. You type a prompt in the web UI
2. The prompt is sent to an OpenCode session via the SDK
3. OpenCode's agent calls browse commands through its bash tool, guided by the browse skill
4. Each action (navigation, click, screenshot, etc.) streams back to the UI as it happens
5. The UI renders a visual timeline: screenshots, summaries, and status updates

• Runtime: Bun
• Backend: Bun.serve() with WebSocket for live updates
• Frontend: React, served via Bun HTML imports
• Agent: OpenCode SDK — manages sessions, tool calling, and event streaming
• Browser automation: browse (Playwright wrapper with persistent daemon)

• Bun >= 1.0
• browse installed and on PATH
• OpenCode installed and configured with an LLM provider

bun install
bun run dev

Early development. Not yet functional.