microsoft/foundry-local ? reverse-engineered prompt

Build me a local AI toolkit that apps can run entirely on the user’s device, with no cloud calls, no API keys, and offline support. I want a lightweight runtime that can download and manage models automatically, pick the best hardware available on the machine, and work with chat completions plus speech to text using Whisper style models. It should feel easy to use from the app side, with simple SDKs and an OpenAI compatible API so existing apps can point at it with minimal changes.

Also include a small local server for sharing models across processes, a command line tool to explore and run models, and a few working samples that show chat, audio transcription, and basic model management. Make it polished enough that a developer could install it and try a model right away. If you need to look up current docs online while building it, go ahead.