jundot/omlx ? reverse-engineered prompt

Build me a server for running LLMs locally on my Mac, optimized for Apple Silicon using MLX.

The main aplication should be a native macOS app with a menu bar icon where I can manage everything. From there, I should be able to start and stop the server and open a web-based admin dashboard. The server itself needs to expose an OpenAI-compatible API, so I can connect other coding tools to it.

Performance is super important, so it needs to have features like continuous batching. I also want a smart, two-tier KV cache that uses RAM as a 'hot' cache and the SSD as a 'cold' cache. This should make inference faster and even let the server remember context across restarts.

The admin dashboard should let me monitor performance, see which models are loaded, and manage them—pinning my favorites, setting timeouts, and unloading others. It should also have a built-in tool for searching and downloading models from HuggingFace. Finally, include a simple chat UI in the dashboard so I can test models, including vision models with image uploads.