Cua Pairs Jev With Driver for Fast, Cheap Computer Use

Cua announced jev-use, a development preview for macOS, Windows and Linux that pairs TypeSafe AI's Jev decision model with Cua Driver, Cua's tool for operating desktop apps and browsers on an agent's behalf. The post opens by declaring that fast computer use is now solved. That is Cua's claim. The opening post gives no speed or cost measurements to support it, and the rest of the thread could not be retrieved for this explanation, so none are reported here.

Cua's longer write-up two days later explains what jev-use is: a public-preview integration recipe with examples. The application reads the current browser state and builds a menu of concrete actions, each with an ID. Jev, which returns a typed decision rather than generated text, picks one ID. The application maps that ID back to the predefined action, checks it, has Driver execute it, then looks at the page again to confirm what happened. A general-purpose model in an agent loop would typically write the tool call and its arguments itself; here they already exist in application code, so Jev cannot invent a tool, a screen coordinate or an argument.

The same write-up is more cautious than the announcement. It says the examples include a mock path that proves the loop is wired correctly but not how the live model behaves, and that a live run needs the user's own TypeSafe API key. It also says faster decisions do not fix missing perception, a page that changed underneath the agent, a poor candidate menu or long-horizon planning, and Jev reads only text, so anything on screen must be described in words first. On that evidence, jev-use shows a constrained, checkable way to move routine choices off a large model, not a finished answer to computer use. Intel covers the write-up, including Cua's own small form model, in a separate item.