I'm working on an iPhone app to run iOS-native agents using both cloud and local models, powered by llama.cpp. It has access basic iOS tools such as calendar, reminders etc. but also more advanced ones like a custom JS environment running on QuickJS that can use various custom modules like an HTTP client, Git etc.
It's a project I have been working on for quite a long time and I released it on TestFlight about a week ago. It was really nice to work on something end-to-end, from creating a wrapper around llama.cpp with support for prompt caching/forking and automatic model loading and unloading based on device memory constraints, to the custom agentic harness the app runs on. I have also spent quite a lot of time on agent execution modes that I hope can enable to more easily reason about agent security regarding prompt injection attacks.
What I'm really hoping for now is to get actual feedback, to know if users end up having real use cases where the app is truly useful / interesting for them, to understand what should most urgently be improved etc.