agent-desktop lets an AI agent click through native Mac, Windows or Linux apps by reading their accessibility tree, not screenshots.
Yes, agent-desktop (Vercel Labs' CLI for AI agents to operate desktop apps) is open source under the Apache-2.0 license.
Free (Apache-2.0), local binary, no paid tier. Optional helper sends screen descriptions to a model provider you configure. (free).
Yes, agent-desktop (Vercel Labs' CLI for AI agents to operate desktop apps) can be self-hosted (the source is available under the Apache-2.0 license).
**agent-desktop** (vercel-labs/agent-desktop) is an **Apache-2.0** Rust CLI (with a C FFI) that lets AI agents operate native desktop applications on macOS, Windows and Linux. It reads each app's **OS accessibility tree** and returns **JSON snapshots with element refs** the agent can act on, so it does not depend on screenshots and pixel coordinates. It is headless by default and exposes 56 operational commands. It comes from the same Vercel Labs team as agent-browser.
Some revenue work still happens in desktop apps with no API: an old on-prem CRM, a finance client, a dialer. agent-desktop gives an agent a structured way to work those apps. Start with read-only tasks such as copying records into a sheet, and keep it away from anything that sends or deletes until you have watched it run.