agent-desktop is an open-source Vercel Labs desktop automation tool, not a model or an autonomous employee. An external agent reads OS accessibility controls as JSON, chooses actions, and checks the UI again. Platform support does not guarantee every app exposes useful controls, or that accessibility always beats screenshots.
The documented minimum loop is install, inspect status and permissions, take a snapshot, act on a real ref from that snapshot, then take another snapshot to verify the result. The README’s macOS example is agent-desktop snapshot --app Finder -i, followed by a click such as agent-desktop click @s8f3k2p9:e8, then another snapshot. The example ref MUST be replaced with a complete ref actually returned by the current snapshot. Finder is a macOS example, not a cross-platform app name. Quote refs in PowerShell. If the UI changes or a ref is stale, observe again rather than blindly repeating an action.
The agent supplies the task loop; the tool offers a CLI, C library, and MCP interface. Set permissions, stop conditions, and confirmation requirements before granting write access. Private screenshots, snapshots, clipboard content, and debug artifacts must not be uploaded indiscriminately. A successful command is not proof of task correctness or authorization to send, delete, or pay.
This is a documentation-based guide reviewed November 7, 2026 (Asia/Shanghai), not our installation test. The original November 3 project roundup is not presented as today’s trending chart. Persistent memory and isolated working directories are separate needs; installing many tools does not replace acceptance testing.