Shipping WREN to the Microsoft Store

On 10 October 2026 I launched WREN, my local dictation app for Windows. Later that night I pulled the launch.

The download worked perfectly on my machine. It did not work on a clean one. This post covers what went wrong, why the Microsoft Store turned out to be the fix, and what I would tell anyone shipping a desktop app on Windows today.

The machine that built it can't see the bug

When you download a file on Windows, it gets tagged as "from the internet". When you unzip it, that tag spreads to every file inside. WREN's zip had 3,472 files in it, and every one of them carried the tag.

On a fresh Windows 11 install, Smart App Control is on by default. It refused to start an unsigned script that came from the internet, and one runtime component refused to load. A buyer would double-click, get nothing, and assume the product was broken.

None of this showed up while I was building it. A development copy of an app never gets downloaded, so it never gets tagged. Every test passed because every test ran on the one machine that couldn't reproduce the problem.

The build machine is the one place a bug like this can't be seen.

So I pulled the campaign that night and wrote two rules that now gate every release:

• Walk the stranger's path. Before anything is offered for download, someone downloads it, double-clicks it and installs it on a machine that did not build it.
• Write the go/no-go down. Before anything goes public: what is proven, what is assumed, what could stop a buyer, and what the worst case is.

Why the Store was the answer

The usual fix is a code-signing certificate. Without one, Windows treats your app as a stranger.

The Microsoft Store offers another route. You package the app as MSIX, and Microsoft signs the package. Smart App Control trusts it, and the install is a real one: Program Files, a Start menu entry, its own icon, pinnable to the taskbar. To the person installing it, it looks and behaves like a product.

It came with a cost: a certification queue, and real people reviewing what you submit.

Three rounds of certification

Round one failed on the privacy policy. My privacy link pointed at a landing page, not a privacy policy. Fair. I wrote a proper policy that separates two things people often blur: the app (which makes no network connection at all) and the website (which counts visits with its own first-party beacon).

Round two failed on wording. The description of the trial did not match the Store's rules about how trials are described. I rewrote it.

Round three passed. WREN went live in the Microsoft Store on 14 October 2026 as "WREN Local AI Dictation", free to install.

The package itself passed every round, including a review of a restricted capability that a human checks by hand. Both rejections were about the text around the app, never the app.

The bugs the packaging found

Packaging also turned up bugs that were hiding in the build I had already shipped.

• A false green on the graphics card. WREN uses a GPU when there is one and falls back to the processor when there isn't. The check that chose between them could pass on a machine that would then fail on the first real sentence, and the fallback was unreachable. The Store build leaves out the large GPU runtime, so it would have failed everywhere. Now the fallback is triggered by a real transcription, not by whether a model loaded.
• The speaking bar never started in one packaged layout, because it was launched from a file that layout didn't contain.

Neither of these showed up in tests run from the source folder. Both showed up as soon as the app was built the way a customer would receive it.

What I'd tell anyone shipping on Windows

1. Test the download, not the folder. Download your own release from a browser onto a clean machine and install it the way a stranger would.
2. Consider the Store even if you sell elsewhere. WREN is free to install from the Store. The $49 licence sells on my own site.
3. Read rejections as notes, not verdicts. Both of mine were fair, specific and quick to fix.
4. Pulling a launch is a decision, not a failure. A first impression you can't install costs more than a delayed one.

WREN is in the Microsoft Store now. The longer story of how it was designed is in the case study.