Agent Desktop

Agent Desktop is an open-source CLI and C library (Apache-2.0, Rust) from Vercel Labs for macOS, Windows, and Linux. It snapshots any app through the OS accessibility tree and returns JSON with a ref on every actionable element. Your agent clicks, types, or sets values on those refs, then snapshots again to verify. Headless by default: no pointer dance and no focus steal unless you pass --headed. npm package agent-desktop ships native binaries; skills wire the loop into coding agents.

Pixel computer-use is brittle on real desktops. If your agent needs Finder, Slack, IDEs, or native apps — not just a browser — you want structured UI trees and actions that fail cleanly when the screen changes, so loops can re-snapshot instead of clicking the wrong thing.

Install with npm install -g agent-desktop, grant Accessibility on macOS if needed, then run agent-desktop snapshot --app Finder -i, act with agent-desktop click @snapshot:ref, and snapshot again. Dense UIs start with --skeleton then drill with --root. Errors carry codes and a retry disposition so agents only retry when safe. Load agent guidance with agent-desktop skills get desktop (and platform), or npx skills add from the repo. Docs and product page: agent-browser.dev/desktop.

Not another browser scraper and not vision-only computer use. Page Agent and Chrome DevTools MCP own the web DOM; Agent Desktop owns native desktop apps via accessibility APIs, with snapshot-scoped refs that go stale on purpose when the UI moves.

• API / SDK surface
• MCP server / client support

• Page Agent ↗
• chrome-devtools-mcp ↗
• Cloudflare OS ↗