A new tool lets AI agents test Windows apps on a clean PC
In a test by its maker, a clean PC found a missing runtime in 29.8 seconds.
A developer released sandbox-mcp version 1.2.0. It starts a new, clean Windows 11 PC in Windows Sandbox, a Windows tool for short test PCs. A person or an AI agent can install an app, use its controls, record the results and remove the PC. The maker gives 40 to 50 seconds for a full test of the screen.
How it works
The tool operates from a command line or as an MCP server. MCP is a standard to connect tools to an AI agent. It first checks the app on the host PC. Then it starts the clean PC with the network off. It records screenshots and logs, and then removes the PC.
The numbers
In one test, an app did not start in 29.8 seconds on a clean PC. A Microsoft C++ runtime was not there. After an install that you select, the app started in 30.9 seconds. The report tells that this result is not from a fully clean PC. The tests include 7 configurations.
What is not known
All results come from the maker. The tool does not work with all Windows apps, and not with apps that use special hardware. For many PCs, data that stays or Linux, the maker tells you to use a VM or a cloud server.
Insight: Agents now need a clean Windows area to work in
Akamai made a sealed area that limits the access of agent code. The sill tool gives agents a Bash shell on Windows, because agents are made for Linux. This tool gives agents a clean Windows PC to test on. Maybe this is a move to control the area where agents work on Windows. That area controls which test results are kept. Look for the tool that your agent tools use first.
Sources
Posted
