Vercel agent-desktop: platform limits and verification checks
Vercel Labs’ agent-desktop gives an AI agent a command-line interface to desktop apps through operating-system accessibility APIs. It provides a CLI and C library; the calling agent still owns the interaction loop. The useful question is not simply whether it supports your operating system, but whether it can observe and verify the particular app you want to automate.
The v1.0.2 release dated October 30, 2026 added the documentation site and changes to the permission guide, among other updates. This is a practical look at the current documentation, not a new launch announcement or a hands-on test.
Check the desktop session, not just the OS
| Platform | Documented requirements and limits |
|---|---|
| macOS | Native x64 and ARM64 packages are available. Accessibility permission is required; screenshots also need Screen Recording permission. |
| Windows | Native x64 and ARM64 packages are available. An app running as administrator can be read, but input requires the agent to have elevated privileges. Ordinary same-integrity-level apps need no separate permission grant. |
| Linux | Ubuntu 24.04 or later with GNOME 46 is the primary target. A graphical user session and AT-SPI accessibility bus are needed. Wayland app-pointer input and native window control are unsupported. |
The overview draws another boundary: agent-desktop reads what apps expose through accessibility, rather than finding controls by matching pixels. A visible control missing from that accessibility tree is not automatically available to it. A screenshot is available separately, but it is not a substitute for accessible controls in the core workflow.
A successful command is not a completed task
The JSON response guide distinguishes action success from evidence of the resulting state. A successful response can mean that an action was dispatched, not that the intended change happened. Even a verified field value does not prove that a document was saved.
For example, a form-filling workflow should distinguish changing a field from confirming the resulting saved record. The snapshot documentation also warns against treating a partial observation as the whole interface: an incomplete snapshot may have hit its observation budget. Element references belong to a particular snapshot, so stale observations need refreshing rather than blind retries.
Review evidence before sharing it
The project FAQ lists the project under Apache-2.0 and identifies screenshots, window titles, clipboard data and debug artifacts as potentially sensitive desktop content. Redacted structured fields do not make a screenshot safe to publish. The native CLI does not require a remote model service, but that does not describe every agent integration: an optional repository helper can send goals and screen descriptions to a configured provider.
AI-generated editorial illustration by ByteForward using the unchanged official Vercel logo. Not a product screenshot or Vercel endorsement. Official brand source. Vercel, the Vercel design, Next.js and related marks, designs and logos are trademarks or registered trademarks of Vercel, Inc. or its affiliates in the US and other countries.
