I've been using the app a bit and I like the idea but one of the major things keeping me from using it is that the model is just fundamentally lazy and without having an easy way to review the code on my phone it's not easy to check its work. If it needs to do something which requires talking to a web API it will often just populate the app with fake data silently. Apps for phones are very dependent on web APIs and with people heavily locking this down these days (when's the last time you've used an API that doesn't ask you for an API token?) I feel like most of the apps are limited to being local-only. There are plenty of nice local-only apps on iPhone, don't get me wrong. I'm sure there is a timer app that is just right for everyone. But most of my tasks require more context about my life, or the outside world. Bitrig isn't able to oneshot this yet–but it definitely does make pretty UIs.

Yeah this is an area we want to make a lot better. What we're currently thinking is giving the model a way to more explicitly call out that it is adding a call to something that needs an API key, which would add something to the UI to allow you to just paste in your key.

I think there's also work when need to do in the system prompt to guide the model away from all of the placeholder data (or at least so that when it adds it there's a better affordance to transform it to real data).