AI learns the job once.Then it runs fast, without second‑guessing.

The clicking, typing, and copying you still do by hand. A person's pointer races through one Monday's chores: signing in to a supplier's portal, downloading invoices, copying a total from a PDF, pasting it into an old desktop accounting app, and sending a summary. The camera pulls back to their computer, where three labels name where that work lives: behind a sign-in, in an app with no API, and in an Excel file.

RPA promised to take it off you. But every robot was built by hand, and fixed by hand whenever a screen changed. A recorded script replays the same chores with a robot's pointer. On a held chord it stops, and a specialist writes each step by hand. The supplier redesigns its sign-in page; the robot clicks where the button used to be, the step fails with "Element not found", and the specialist comes back to fix it.

Now, AI builds it. The three places, checked off, above the headline: AI builds the automation. Your computer runs it.

Describe the job, and Kitewell lays out every step for you to check. In Kitewell's chat window, the request is typed: “Every weekday at 9, collect new invoices from the Harbor Supply portal, enter them in Ledger, and email me a summary.” Kitewell proposes the workflow as a card with its steps drawn as a graph; applied, the workflow draws itself in the editor.

Or build it from Claude Code or Codex, over MCP. For developers: the workflow's YAML, then a terminal where a coding agent changes it through Kitewell's MCP tools, checking the draft and then saving it.

On the first run, AI learns the job in your browser. Then your computer runs it, every day. Real Kitewell runs against a stand-in supplier portal on the same computer. The first run's record, each action worked out as it goes, beside what the browser saw; then the next run, its sign-in actions marked Replayed.

When a site changes, AI finds the new way, and keeps it. The portal moves its sign-in button. In the next run's record, "Click Sign in" is marked Recovered: AI found the new button. In the run after that, the same action is Replayed.

It stops for your OK where it matters. A notification says a step is waiting for the person. One click approves it, and the run's record shows every step done and who approved it.

Got a list in Excel? It runs every row and writes the results back. A sheet linked to orders.xlsx runs the workflow once per row; then the workbook itself fills in each row's result.

Coding agents too. Claude Code writes the fix, Codex reviews it, and nothing ships until you approve. A real workflow on a demo repository: Claude Code fixes a failing test, with Codex and then Gemini behind it if it fails; Codex reviews the change; the run waits at the approval with Approve and Send back before the fix is committed on a local branch.

It runs on your computer, and your passwords stay there. On another computer that took the project, its portal password named but not filled, the person there sets the value, and the page notes that values stay on each device.

Run it on every computer, and watch them all from the cloud. If a run fails, or a computer goes quiet, you hear about it. The two computers report their runs to Dagu Console, whose real computers page takes the frame. Kitewell's alert for a failed run arrives in the team chat, and Dagu Console's email says a computer stopped reporting. Notes say Kitewell's alerts need Personal or Team, and Dagu Console's stopped-reporting alerts are turned on per computer.

Kitewell. It flies on its own. You hold the string. The kite leaves the logo and flies, its string running back to the last word. kitewell.app. Free for one project, on macOS and Windows.