Most useful open-weight models still don’t have a hosted API. You either rent a GPU, leave a Mac running overnight, or send prompts somewhere that can read them. Umbra is my experiment: an OpenAI-compatible endpoint (Anthropic-compatible too) where Apple Silicon owners host public Hugging Face GGUFs they approve. You swap base_url, keep your existing client, pay per token. Try it: https://tryumbra.dev Quickstart: https://tryumbra.dev/docs/quickstart Host a Mac: https://tryumbra.dev/setup

Hey Product Hunt — I’m Thomas. I built Umbra. I kept hitting the same wall: a model I wanted only existed as a GGUF on Hugging Face. No API. Options were rent a GPU, self-host, or trust a random endpoint with my prompts. So I built a boring interface for that. Swap base_url, mint a key, hit Macs that are code-attested. Privacy isn’t “we pinky-promise we don’t log.” The design is: decrypt in-process, don’t persist, fail closed if trust checks fail. Honest alpha limits: two live models, HK-only cash-out for providers, no production SLA. I’m not pretending this is Together with a different logo. I’ll be in the comments all day. If you try it, I’d rather hear: does the API behave like a real OpenAI-compatible provider under stress? is the trust story clear or hand-wavy? which public GGUF should be next? Quickstart: https://tryumbra.dev/docs/quicks...

the OpenAI and Anthropic compat layer is a nice touch, means i can swap base_url and actually use my existing client instead of writing a wrapper. love that you sidestepped the GPU rental trap

Umbra

OpenAI-compatible API for models nobody else hosts

OpenAI-compatible API for models nobody else hosts