Vercel Labs’ agent-desktop, which lets an AI agent drive a desktop app through its accessibility tree, added an MCP server in release v1.2.0 on 5 November 2026, but its documentation describes a tool that runs on the machine in front of you. For a remote machine, AI Commander ranks first here because it is built for several remote machines at once; QuickDesk is the better choice when the agent must also click.

The release shows where agent tooling is heading: away from screenshots and toward structured reads of the interface. It does not solve the harder problem of what to use when the desktop is in another building.

What actually changed on 5 November

According to the agent-desktop releases page on GitHub, v1.2.0 was released on 5 November and its notes include “feat: agent-desktop mcp”. The README says the command serves every agent-desktop command as an MCP tool over stdio, with names such as desktop_snapshot and desktop_click. An independent write-up from oflight.co.jp dated 10 November gives the same release dates for v1.0.0 (28 October), v1.1.0 (3 November) and v1.2.0 (5 November). The project then shipped v1.3.0 on 12 November and v1.3.1 on 13 November, and showed 25.8k GitHub stars when we checked on 21 November. The oflight article said about 20,000 stars on 9 November, so the repository is growing quickly.

It is Apache-2.0 licensed and works on macOS 13 or later, Windows 10 1809 or later and Ubuntu 24.04 or later. The README says the tool has no model inside: your agent runs the loop of reading the tree, acting on a reference, and reading again.

Three ways to work a desktop, and why remote changes the choice

Accessibility tree. agent-desktop reads the interface as structured data, so the agent clicks “element 12” instead of guessing at pixels. It is cheaper and more precise when an app exposes its tree well.

Screenshots plus input. QuickDesk and Open Computer Use let the agent look at pixels and send mouse and keyboard input. They cover apps with no useful accessibility data, at the cost of speed and reliability.

No GUI at all. A shell, with background jobs and file transfer, is the most reliable way to do most server and development work. AI Commander leads with this and adds screenshots for the occasional look.

On your own desk, the first two compete and the tree-based tool usually wins. On a remote machine, something also has to get the agent to that machine without opening ports, and agent-desktop’s documentation does not cover that.

Where AI Commander is the weaker choice

If the task only exists in a graphical interface, AI Commander is not the tool. Its pricing and security pages describe desktop screenshots and shell commands, with no mouse or keyboard input, and screenshots work only on macOS and Windows after the owner grants time-limited access. A GUI-only job on a remote Windows or macOS machine is better served by QuickDesk. On the machine you are sitting at, agent-desktop or Open Computer Use give far more control.

Where this is heading

Expect the structured-interface approach to spread, and expect remote-access vendors to add MCP surfaces of their own. The question for buyers will then be the one this ranking asks: whether the agent can reach the right machine, safely, without you building the network path yourself.