Open-source Jev-cu makes Codex computer use faster by passing text, not screenshots
alexcovo_eth · x · 2026-09-19
Developer Sacccc open-sourced Jev-cu (138 stars on GitHub), which offloads the "where to click next" decision in computer use to Jev (TypeSafe System One): Jev picks the element, action, completion and risk level from on-screen text candidates, while Codex reads the UI and executes, with a local policy gate blocking sensitive operations.
Key points:
- Passes text only (AX snapshots), no screenshots, making it faster and smoother than Codex's built-in computer use at similar token cost
- Ships an installable Codex skill (npm run install-skill into /.codex/skills/)
- After setting a Typesafe API key, tasks run via a runTask loop in Codex's cuarepl runtime, with dryRun and maxSteps options
- A Mac Calendar demo shows the Jev version running nearly pause-free; speedups are biggest on tool-heavy tasks
- Includes unit tests and offline evals measuring element-selection accuracy on AX snapshots
Related event: Open-Source Jev-cu Gives Codex Faster Computer Use Capabilities(2 posts)→
More from coding & agent
- Grok Build tests Remote Control: link your PC via CLI and drive it from your phone — XFreeze · 2026-09-19
- Developer Says AI Models Now Handle Nearly All His Work — He Just Approves — tushaarmehtaa · 2026-09-19
- 8 components of harness engineering: why the system around the model, not the model, makes agents reliable — blaizedsouza · 2026-09-19
- Google's 12-page Agentic Engineering guide lays out a 5-stage pipeline for building agent teams — blaizedsouza · 2026-09-19
- AgentRun: a purpose-built harness for high-frequency repetitive knowledge work — blaizedsouza · 2026-09-19
- Qwen 3.8 27B on a single RTX 5090 builds a full animation using only code — Acceptable-Object390 · 2026-09-19