# agent-desktop

> Apache-2.0 desktop computer-use CLI — OS a11y refs, not pixels

**Category:** AI Agents  
**Pricing:** open-source  
**URL:** https://github.com/lahfir/agent-desktop  
**GitHub:** https://github.com/lahfir/agent-desktop  
**Added:** 2026-09-30

---

## The Hook

agent-desktop gives agents real desktop computer use through OS accessibility trees — stable refs and retry-safe actions instead of screenshot guessing.

## What's Good

Apache-2.0 (lahfir/agent-desktop). Rust CLI + C-ABI cdylib; progressive skeleton snapshots (claimed ~78–96% token cuts on dense UIs); ref-based click/type; headless-by-default a11y actions; CDP launch path so Playwright/agent-browser can drive Chromium apps while native chrome stays on a11y. npm `agent-desktop` / npx; ClawHub + skills.sh listings.

## The Catch

Production a11y drive is macOS-first — Windows/Linux a11y and input still Planned in the support matrix (cdylib ships wider). Vs agent-browser (Vercel Labs browser CDP/a11y refs): agent-browser is the web/browser CLI; agent-desktop is native desktop apps via OS accessibility. Needs Accessibility (+ Screen Recording/Automation for some features); do not hand it a logged-in machine without a sandbox story.

## Verdict

Apache desktop computer-use when agents must drive native macOS apps by a11y refs — not agent-browser’s web CDP lane; Win/Linux still catching up.
