Kindred Coach

A private AI coaching app that runs entirely on your device. Pick a coach, have a real conversation, and get clear takeaways - no accounts, no cloud, no tracking. Just you and your growth.

Comment

Inspiration

I wanted coaching — not therapy, not a chatbot, just someone to help me think clearly. Good coaching is expensive and hard to access, and most AI apps send your private thoughts to a server somewhere. When Apple announced Foundation Models at WWDC, I saw the chance to build something different: real coaching conversations that never leave your device.

What it does

Kindred Coach gives you 5 expert-designed AI coaches — each with a unique personality and proven coaching framework. You pick a coach, have a natural conversation, and get auto-generated insights at the end: key themes, patterns in your thinking, and a concrete next step. Coaches remember you across sessions, so the more you talk, the more personal it gets. You can also create your own custom coaches and share them with friends.

How we built it

SwiftUI and SwiftData for a fully native iOS experience. Apple Foundation Models for on-device AI — no network calls for conversations. Structured generation using @Generable for consistent session insights. RevenueCat for a simple one-time Pro upgrade. The prompt system has two paths - a rich template path for structured coaches and a fallback path for custom coaches created by users.

Challenges we ran into

The on-device model has a roughly 4,096 token context window. Longer conversations would fail, so I built session rotation — the app summarizes the conversation so far and starts a fresh AI session seamlessly. The model also occasionally leaks its own system prompt into responses, so I added a real-time sanitization layer that strips leaked content during streaming. Swift 6 strict concurrency made passing SwiftData models across actor boundaries tricky — I had to restructure APIs to pass primitive values instead.

Accomplishments that we're proud of

The app genuinely feels like talking to a coach, not a chatbot. Coach memory works — starting a second session and hearing the coach reference what you talked about last time is a small moment that feels like magic. The entire experience is private by default, not as an afterthought.

What we learned

Privacy doesn't have to mean compromise. Running AI on-device creates a better experience — people open up more when they know no one is listening. On-device models have real constraints, but working within them forced creative solutions that made the product better. The on-device model works best with focused, conversational coaching. Complex multi-step planning can hit the model's limits today, but Apple Intelligence capabilities will keep improving with each generation of hardware and software — and Kindred is built to take full advantage as they do.

What's next for Kindred Coach

iPad support with a multi-column layout. Richer coaching frameworks with guided exercises. Trends and patterns across sessions over time. And exploring group coaching - sharing progress with a trusted friend or accountability partner while keeping conversations private.

Built With

Share this project:

Updates