Windows Sandbox in 30 Seconds: sbx and an MCP Server for One-Off Tests

⚡ The gist in 5 seconds

• Scott Hanselman has released sandbox-mcp v1.2.0 — a CLI sbx and an optional MCP server that spin up a disposable Windows 11 environment in Windows Sandbox.
• Installs without Node/npm: the portable build already ships with Node.js, a guest agent, and setup instructions.
• The guest is clean every time and offline by default, while dependency installation only happens with the explicit --provision flag.

🔍 What happened

The shanselman/sandbox-mcp repository has racked up 58 commits and matured into the v1.2.0 release: a portable CLI plus an optional MCP server, bundled Node.js, a guest agent, and setup instructions. The primary interface is the sbx command. The MCP adapter is only needed for clients that want the described tools and images right in their responses; both interfaces run the same engine and share state.

The mechanics are simple. sbx deps app.exe finds what the application is missing, and sbx run app.exe --provision delivers the runtimes — but only when explicitly asked. The README includes a measured walkthrough: NativeProof.exe on a fresh Sandbox fails in 29.8 s with code -1073741515 (missing VC++ redistributable), and after provisioning it PASSES in 30.9 s. The installer is Microsoft-signed 14.51.36247.0, downloaded and verified on the host, with the guest's network switched off the whole time. The guest can be driven via UI Automation: inspect, type, click, toggle, wait. Output includes PNGs, length-capped recordings, self-contained HTML, and optionally MP4. Frameworks supported: WinUI, WPF, WinForms, Win32, MSIX/AUMID activation. GUI samples in the README take roughly 40–50 seconds from machine start to teardown.

🧠 Our take

Our assessment: this release took off not because it does anything technically impossible, but because MCP has become the socket through which agents reach developer tooling. And Hanselman is no anonymous author — he has years of habit of publishing "tools built for himself," and you can see the same logic here: first a CLI that works on its own, then an adapter.

What this actually changes. "Works on my machine" stops being a joke and becomes a command: a clean guest with no years of accumulated junk catches a missing dependency in half a minute. Networking is off by default, and the feature list includes installer diffing and "detonation" monitoring — meaning a suspicious build can be dissected without risking your work laptop. We believe the "clean + offline + UI automation" combination is what's truly valuable here, more than MCP itself.

Who wins: Windows desktop app developers and QA folks tired of manually reinstalling the system. Who loses: commercial device farms for testing and checklists where "we tested it on our machine" is the only artifact. One caveat: Windows Sandbox, in my experience, lives in Pro and Enterprise editions, but I haven't verified the edition requirement in the README.

⚠️ What we don't know

There's a single author, and the README is its own press release. The 29.8 / 30.9 s and 40–50 s figures are backed by nothing external: no independent runs exist. There's no data on RAM and disk usage, on ARM64, on parallel runs, or on what happens to the project if Hanselman loses interest. I've only seen the license as a LICENSE file — I haven't read the text. Verifying it yourself is easy: run sbx deps on your own app and compare against what's promised.

Prediction: within a quarter, at least one fork or wrapper around sbx will appear with parallel execution of multiple Sandboxes — the README says nothing about it, and that's exactly what real CI will demand.

Disclaimer: The 29.8 / 30.9 s figures, the 40–50 s GUI sample time, and the entire list of supported frameworks come solely from one author's README and have no independent runs — there are no third-party benchmarks or audits of the project.

Source: github.com/shanselman/sandbox-mcp