How can Jev use a computer?
Published
Jev is an AI model that selects from supplied options. A computer-use application gathers screen information, asks for a choice, checks it and executes the corresponding action.
Explanation & code
This is based on a Mac computer-use project. Our JavaScript companion uses the real SDK with a simulated answer and fake screen, not live computer control. A valid choice can still be wrong.
Understand it. Then fix it.
Read the screen and define actions
The referenced Mac project reads screen text and accessibility controls. Our small demo offers Save, Cancel and Wait for the goal “Save this note.” The screen and response are fixtures.
Ask for a choice
The JavaScript SDK sends the goal, screen and options. The returned choice is data; it has not clicked anything.
const response = await client.systemOne({
state: { goal, screen },
questions: { next: choice("Next step?", options) }
});Check, execute, verify
These are our application helpers. Map only known choices to actions, check permission and reread the result after execution. In the demo, the click counter changes from zero to one and the fake note becomes saved.
const action = actions[answer.choice];
if (!action || !allowed(action)) return;
await action.run();
await checkSaved();Keep the limits visible
Stale screens, missing controls and wrong choices can break the flow. Require approval for risky actions. The referenced project uses a separate text model to write messages; a fixed workflow may only need a normal script.
Code blocks are teaching excerpts. Keep the surrounding error handling and application requirements.
Save the code excerpts ↓Read the full transcript
Jev is an AI model that picks from options. How can it use a computer? In this Mac project, software reads screen text and the controls an app exposes. It gives Jev a goal and possible actions. Our JavaScript demo asks: save this note. The options are Save, Cancel, and Wait. Jev returns a choice. Our answer is simulated. So the answer says save. Has anything clicked yet? No. Our code maps save to the button. It checks permission, then clicks. Read the screen again to check that the note was saved. A choice can still be wrong. Missing controls or moved buttons can break the flow. Use fresh screen data and keep risky actions behind approval. And if it needs to write a message? This project uses a separate text model. For fixed steps, a plain script may be enough. Jev helps when the next action needs judgment. Jev picked Save. Your code would like some credit.
