agent-desktop
Computer use for coding agents from Vercel Labs: a command-line tool (with a matching agent skill) that reads desktop apps on macOS, Windows and Linux through the operating system's accessibility tree, gives every button and field a reference, and lets the agent click or type on that reference and check again, with no screenshot matching. macOS needs Accessibility permission; Linux needs a desktop session with at-spi2-core. Snapshots can contain private screen content. v1.0.2 released Oct 30, 2026.
Test status
Not tested yet We list this skill from its source; we haven’t run it in a client yet, so there is no verdict.
Get it
The skill’s files and install instructions are in its repository: github.com/vercel-labs/agent-desktop.