agent-desktop is an open-source CLI and C library from Vercel Labs for coding agents that need to operate desktop apps on macOS, Windows, and Linux. A snapshot reads the target app through the operating system accessibility tree and returns JSON with a ref for each element the agent can act on. The agent clicks, types, or toggles that ref, then snapshots again to check the result. Screenshots are a separate command. The main loop does not require them or pixel matching. Vercel Labs already ships agent-browser for structured browser control. This release applies that approach to native apps, including menus, forms, and end-to-end tests of desktop software. Ref actions are headless by default and go through the accessibility API, so they do not move the pointer or take focus while you keep working. Several agents can use one computer at the same time. The launch says each agent gets its own cursor, and the CLI includes a cursor-overlay command. On a crowded window, a skeleton snapshot returns a shallow map, and a later call can drill into one region so the agent reads only the part it needs. Lahfir built the project in Rust and had just joined the Vercel Labs team with this release. Chris Tate, who works on that team and created agent-browser, posted the announcement. You install it with npm install -g agent-desktop. The binary includes skill docs an agent can load before the first command, and the code is Apache-2.0.