From bug report to proof,on one canvas.
The agent starts the dev server in a terminal card, finds its port on the Dev servers card, opens the app in a browser card and reproduces the bug the way a user would. Then a failing test, the fix, green tests — and the same clicks in the browser, this time working.
A coupon turns the cart total into NaN. Answer the agent when it asks to run the tests.
Claude Code starts the dev server, finds its port, reproduces the bug in a browser card, pins it with a test, fixes it and proves it.
Reproduced, not guessed
The agent clicks through the real app in a Chrome card and reads the page back with browser_snapshot before it touches the code.
A test that pins the bug
It writes the regression test first and runs it in a terminal card you can watch — red before the fix, green after.
The real port
The Dev servers card lists what is listening; ports_list tells the agent where its server actually came up.
You hear when it is done
“Needs your input” and “finished” are separate notifications — on the desktop, on your phone or in Telegram.
The loop, card by card
Five cards and the arrows between them.
More ways to ship with agents
Every scenario runs on the cards you already saw — pick the next one.
Give your next bug to an agent
Your agents, your keys, your machine.
macOS and Linux are coming later.
The first-run wizard installs an agent CLI, Git and the dictation model for you — or skips each step.