Free open-source method has AI click through your web app to visually test it
A built this to solve a common problem: spending hours manually testing different app flows and edge cases before launch, especially since bugs from AI-assisted ("vibe") coding are easy to miss. Claude Code alone was tried first, but it moved too slowly and didn't track what it had already done, what flow it followed, or where it had been, so it behaved mechanically like a bot rather than intelligently. The fix is a "one brain, one call per turn" design: a single continuously reviews where it has already gone, decides the next action, and logs its reasoning, the action taken, and the location into a file that feeds the next turn.
For actually seeing and interpreting the screen, replaces Claude because it's much faster while still being capable enough for this visual-analysis role. Every turn outputs strict JSON covering the screen state, action, and reasoning, allowing the system to walk through flows, edge cases, and full s, and it draws es around visual inconsistencies it spots on screen. The entire technique is released fully open source and free to plug into any agent.
It can even be wired into Claude Code directly to build a custom testing skill for a web app, though that comes at the cost of Claude Code's slower speed.
Key points
- Claude Code alone was too slow and didn't track prior screens or flows well
- A 'one brain, one call per turn' loop keeps reasoning and actions consistent across turns
- handles fast visual perception while a separate model handles reasoning
- Each turn logs screen state, action, and reasoning as strict JSON for the next turn to use
- Covers , edge cases, and full flows, drawing es on visual bugs