A clearer view of AI.
THE INDEPENDENT AI.

AGENT INFRASTRUCTURE · WORTH A LOOK

agent-desktop

Free Vercel Labs tool that lets AI coding agents see and operate desktop apps on Mac, Windows and Linux.

Visit github.com ↗

What agent-desktop does

Open-source computer-use CLI (and C library) from Vercel Labs, the desktop sibling of agent-browser. It lets any coding agent (Claude Code, Codex, Cursor and others) read and operate desktop apps on macOS, Windows and Linux through the operating system's accessibility tree: a snapshot returns a stable reference for every clickable element, the agent acts on it, then checks again, with no guessing from screenshots. Headless by default (no mouse movement or focus stealing), with optional on-screen cursors, one per agent. Install with npm install -g agent-desktop.

Good for

  • Let a coding agent click through a desktop-only app (accounting, CRM or design software) that has no API
  • Have an agent fill in or check the same form in a Windows or Mac app every week
  • Test your own desktop or Electron app end to end with an AI agent

Try this in 10 minutes

  1. Install and grant access. Run npm install -g agent-desktop, then agent-desktop permissions --guide on a Mac (Accessibility permission).
  2. Try one safe app. Ask your coding agent to snapshot a harmless app such as Finder or Notepad and click one item, then check the result.
  3. Pick one real chore. Choose one repetitive task in a desktop-only app and note what the agent needed fixing before you would trust it unattended.

Mini-tasks based on common small-business uses for this tool.

Honest take

Some claims not independently checked

  • Real and free: Apache-2.0 on GitHub (vercel-labs/agent-desktop, about 4.1K stars on 31 Oct 2026), installs with one npm command.
  • Not brand new: the repo dates from Feb 2026; version 1.0, which added Linux, shipped on 27 Oct 2026.
  • 'Agents in parallel, each with its own cursor' is partly true: the cursors are an on-screen overlay per agent, and real mouse and keyboard input is still taken one agent at a time.
  • Only as good as each app's accessibility data; macOS needs Accessibility permission and Linux support is weaker on Wayland.
  • Snapshots can capture private screen content. It is a Vercel Labs experiment, not a supported Vercel product, and needs a coding agent to drive it.

Pricing

Free, open source (Apache-2.0). You pay only for the AI agent or model you run it with.

Prices as listed by the maker when we last checked (2026-11-01); usually in US dollars. Check the website for current prices.

Latest update

  • : Video notes: the 20-second demo shows labelled cursors in Notes, Reminders and Calculator on a Mac, Event Viewer on Windows 11, and a terminal agent reading controls and clicking by element id. It can't prove real input runs in parallel (the repo says one agent at a time), and Linux isn't shown.
  • : Added: Chris Tate (Vercel Labs) announced agent-desktop, 'agent-browser for your whole computer': computer use for any coding agent on macOS, Windows and Linux (v1.0 shipped 27 Oct, v1.0.2 on 30 Oct).

Is agent-desktop worth it? Quick answers

Is agent-desktop worth it?

Worth a look if you or a developer already use a coding agent and have desktop-only software you want automated. It is free and open source, so the main cost is setup time; non-technical teams should wait for it to appear inside the agent apps they already use.

How much does agent-desktop cost?

Free, open source (Apache-2.0). You pay only for the AI agent or model you run it with.

What is agent-desktop used for?

Letting AI coding agents such as Claude Code, Codex or Cursor see and click through desktop apps on Mac, Windows and Linux, for example apps with no API.