COMMUNITY OBSERVEDMetric claims: UNVERIFIEDSOURCE AUDITEDBrowser Agents

Jev Use - Mac Calendar Automation Benchmark

10x faster macOS computer use with Codex and Jev eliminating LLM thinking pauses.

Overview & result

Sac created 'Jev Use', combining Codex with TypeSafe's Jev model for desktop computer use on macOS. In a side-by-side benchmark creating Mac Calendar events, Jev handles UI element selection and click targeting in single-digit milliseconds while reserving heavy models only for text input, resulting in an estimated 10x speedup with near-zero pauses and identical token consumption.

Show the original post (Chinese)

卧槽,居然真有比Codex内置computer use更快的操作电脑的方式

我尝试用Codex+Jev打造了一个加强版computer use,我称之为「Jev Use」。比内置的更快更丝滑,token消耗却差不多

下面是我用「添加Mac日历事件」做了一个对比视频,同样的内容,明显可以看到Jev版本整体过程几乎无任何停顿😆

— Sac

How Jev fits in the loop

  1. Codex reads the macOS screen state and accessibility element hierarchy.
  2. Jev selects the target calendar field or button in single-digit milliseconds.
  3. Codex executes the click immediately and invokes generative completion only when entering freeform text.

How to reproduce

  1. Clone Codex CLI computer use harness on macOS with Accessibility permissions enabled.
  2. Configure Jev as the decision layer to predict UI element coordinates from the system accessibility tree.
  3. Run benchmark: Add Calendar event 'Meeting with Alex' at 3pm Friday, comparing native Codex vs Jev Use.

Why this build matters

Demonstrates that desktop computer use latency can be solved by decoupling fast System One UI action choices from slow generative LLM reasoning.

Limitations

  • Requires macOS Accessibility API permissions.
  • Non-native web canvases without accessibility nodes require visual OCR fallback.

Patterns