Every frontend developer with an AI coding assistant knows this dance:
Notice a misaligned button in Chrome.
Take a 4 MB Retina screenshot (~1,500+ vision tokens).
Alt+Tab to your editor, paste the image, and type: "Move the second button on the checkout card a bit to the left."
Watch the agent guess the wrong file, edit the wrong wrapper, or ask for the console error you forgot to copy.
When an agent is blind to your running tab, you become its human clipboard.
I built Jolt ( vite-plugin-jolt ) to kill that copy-paste loop. Instead of throwing pixels at a vision model, Jolt injects a tiny dev-only bridge into your real browser tab and sends ~280 tokens of exact runtime truth to your editor.
“Stop paying vision-model taxes to point at a CSS flex container.”
Act 1: The Two Dead-Ends of Frontend AI Debugging
Today, you basically have two ways to let an AI agent inspect your UI:
Manual screenshots + pasting logs: Slow, expensive in tokens, and blind to React/Vue component trees, store state, or…