Vercel Labs releases agent-desktop CLI for AI agents to control desktop apps via accessibility trees
·2026.11.11 09:30
Key point
The tool supports macOS, Windows, and Linux by reading OS accessibility APIs to return JSON structures with unique refs for agent interaction, avoiding pixel-based screen matching.
Details
- Vercel Labs released agent-desktop, a CLI and C library that enables AI agents to manipulate desktop applications on macOS, Windows, and Linux by reading OS accessibility trees rather than using screenshots or pixel matching.
- The system assigns unique refs to UI elements within a JSON snapshot; agents interact via these refs in a loop, with the tool rejecting stale refs or ambiguous targets when the UI changes.
- It operates headlessly by default using accessibility APIs without moving the pointer, but supports a --headed flag for physical input simulation.
- The project includes 60 commands (56 operational), MCP support via stdio, and a C library (libagent_desktop_ffi) for in-process integration, licensed under Apache-2.0.
Platform Requirements
- macOS: Requires macOS 13+ (Intel/Apple Silicon) using the AX API.
- Windows: Requires Windows 10 1809+ (x64/ARM64) using UI Automation.
- Linux: Requires Ubuntu 24.04+ (x86_64/aarch64) using AT-SPI on X11 or Wayland.
Integration and Security
- Agents can retrieve usage guides via agent-desktop skills get desktop and integrate through MCP tools like desktop_snapshot and desktop_click.
- Security measures include piping text via --stdin for sensitive commands to prevent process argument exposure and warnings against logging snapshots or clipboard data due to potential PII.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.