It reads your repository before it touches it. Then it runs your app, clicks through it like a tester, follows production errors back to the line that caused them, and hands you the fix to keep or undo.
Your app runs in a sandbox, in a browser tab inside the editor. Select the thing that's wrong, say what you want, and kino edits the source, reloads and checks the same screen.
Pair with kino and the editor splits: you on the left, kino on the right. Its edits land in the open files as you watch, it sees your unsaved typing, and it keeps off the lines you're working on.
kino starts the app, walks every page on desktop and phone screens, types into the forms and submits them. It replays each problem before reporting it, so you get a bug with steps, a screenshot and the line of code.
When Sentry reports something new, kino finds the user's PostHog session, sees what they did, reads the logs of every service around that second and opens your code. You start from the cause, not the stack trace.
src/pages/Cart.jsx:10 stores data.items without checking r.ok; after a 401 it's undefined.kino/fix-shop-1a, ticket linked, a PR with only the fix's files.kino draws your architecture from the code, every box pointing at the file and line it came from. Draw what should exist and kino plans it, you approve, and it builds it, databases and queues included. Every branch and pull request is then checked against the drawing and against the bugs your team already fixed once.
⌘L adds your selection, @ adds files, folders and problems.kino arch mcp gives any agent your drawing and its drift.npx kino-cli, with ask, auto-edit, plan, research and auto modes.How it was measured. Eight coding tickets (two features, a bug and five accounting-domain tasks) on the same small Python repo, one attempt each, November 2026. Both agents ran claude-opus-5-5 at high effort through the same Claude Code login, got the same ticket and conventions, and had to push a branch and open a real PR. A ticket counts as resolved when hidden tests, written before the runs and never shown to either agent, pass on the PR. Time is the median, tokens and cost the mean (cost at API list price). Both resolved every ticket, so this doesn't show either one is more accurate, and eight tickets is a small sample. The harness ships with kino, so you can rerun it on your own tasks.
npx kino-clibrew install dogukangun/tap/kino-cli