LabscoMCP early access ↗
REPO PACKAGECOMMUNITY

callstackincubator ai

callstackincubator · publisher1,360 GitHub starsMIT · Listed as freegithub.com/callstackincubator/ai

LABSCO SUMMARY

Summary

callstackincubator/ai ships exactly one skill, and it is a set of integration recipes for wiring React Native AI's three on-device model backends into an app, not a new capability of its own.

The underlying package, React Native AI, gives a React Native app three interchangeable on-device backends behind the Vercel AI SDK's API: Apple's Foundation Models (built in on iOS, no download, but text generation needs iOS 26+ with Apple Intelligence), Llama.rn (any GGUF model, downloaded from Hugging Face at runtime), and MLC LLM (a fixed list of prebuilt models — Llama 3.2, Phi-3-mini, Mistral-7B, Qwen2.5 — needing 1-8GB of device memory and Xcode's Increased Memory Limit capability). All three cover text generation; embeddings, transcription, and speech synthesis are Apple-only.

The skill itself does not run inference — it is a recipe set that tells an agent which package to install and how to call generateText, embed, transcribe, or speech for whichever backend you pick. It is for a React Native developer adding on-device AI who wants privacy and no server cost, and who is willing to choose a backend per platform and per model size rather than getting one that works everywhere unmodified.

READ MORE

What it costs to start.

Apple's backend needs nothing beyond an iOS 26+ device with Apple Intelligence enabled for generation (iOS 17+ covers embeddings only). Llama and MLC both require downloading a model — MLC's four listed models run from about 1GB to 4.5GB — and MLC's own README warns that any model outside that list means recompiling its runtime from source.