DEMO / RECORDING
Overview & result
A clean open-source desktop control agent for macOS that leverages Jev's high throughput to make real-time interaction judgments per second.
From the original post
Built a macOS computer use agent with Jev. One sub-50ms typed decision per step for mouse, keyboard, and app actions instead of heavy multimodal model prompts. ~$0.0002 per step.
How Jev fits in the loop
- Inspect native macOS UI hierarchy.
- Query Jev Choice with target goal and available UI actions.
- Execute selected OS event and verify state transition.
How to reproduce
- Inspect the source implementation at https://github.com/awlevin/typesafe-computer-use.
- Configure local Jev API credentials and register required tool/model bindings.
- Execute test prompts and observe decision latency and routing outputs.
Why this build matters
Shows that OS agent loops can run at sub-cent budgets with deterministic safety.
Reported performance
Reported by authorCost: ~$0.0002 / step
Author reported cost of approximately $0.0002 per decision step.
Limitations
- The author's caveat: reasoning a frontier model does for free, such as reading dates off the screen, has to be rebuilt here as deterministic state.
- Needs macOS 14 or newer with Screen Recording and Accessibility permissions; Windows support is experimental.
- The loop stops itself on low confidence, a stall or the step limit.