Why are they tying the client with the AI model instead of using OpenAI or another popular http format? Running llama.cpp locally is about the right level of complexity for most doing local AI.