sandbox-mcp: throwaway Windows machines for agents

Summary

sandbox-mcp gives you or an agent a fresh, disposable Windows 11 machine through Windows Sandbox. It deploys an app, drives its UI, records crashes and evidence, and then throws the machine away. It has a CLI (sbx) and an optional MCP server that share one engine. It is MIT-licensed, at v1.2.0, and gained about 340 GitHub stars in its first three days (GitHub trends, 2026-11-09). date_published is the approximate date of first appearance.

Key points

Good for: checking that an app runs on a clean PC without your dev runtimes, testing installers, and opening untrusted files safely ("detonation"). A GUI run takes about 40–50 seconds end to end, by the author's measurement.
Needs: Windows 11 24H2 or later, Pro/Enterprise/Education (not Home), with virtualization on and at least 4 GB RAM. It does not run on Linux or macOS.
Install: download the portable ZIP from the latest release (it bundles Node.js), check its .sha256, then run .\sbx.cmd doctor. If the feature is off, run Enable-WindowsOptionalFeature -Online -FeatureName Containers-DisposableClientVM -All as admin and restart. Open Windows Sandbox once to install wsb. Try .\sbx.cmd run C:\Windows\System32\charmap.exe --open.
MCP: add a stdio server whose command is the bundled node.exe and whose argument is src\server.js (VS Code .vscode\mcp.json; the README shows the Claude Desktop and Copilot CLI shapes). Copilot CLI can also install it as a plugin: copilot plugin marketplace add shanselman/sandbox-mcp.
Tools: sandbox_run (boot, deploy, launch, observe, collect, destroy in one call), lifecycle tools, file copy and exec, UI tools (sandbox_screenshot, sandbox_click, sandbox_ui_tree, sandbox_ui_action), recording, sandbox_install_test, sandbox_detonate, and sandbox_snapshot / sandbox_diff for change tracking.

Why it matters

Agents get a real, throwaway desktop to test Windows software in, so they can check their own work end to end without touching the host machine.