Early access · Paid API credit is live. New accounts still get free starter credit.
PricingDocsModelsFAQEarnBlogAboutSign in →

Home · about

About

Idle Macs, private prompts, open models.

Umbra is a marketplace for private-prompt inference on attested Apple Silicon: it puts idle Macs to work and gives developers a home for the models incumbents will not host.

Machine owners host public models on hardware they already own and earn per token; developers reach them through an API that speaks the formats they already use. The hard part is trust, so prompt privacy is designed as a property of the system, not a policy: an adversarial provider should not be able to read your prompt. Apple hardware and APNs running-code attestation are live for admitted providers; signed per-response receipts are still pending.

That is the name. In a total eclipse, the umbra is the innermost shadow: the region where the light source is completely hidden from view. Your prompts run inside it, and attestation is what makes the darkness provable instead of promised — hardware and running-code identity are live; signed per-response receipts are still rolling out during alpha.

For machine owners

Monetize idle Apple Silicon

Turn idle Macs into an inference provider that hosts public open-weight models and earns per token, on hardware you already own.

For developers

The long-tail models incumbents will not host

Uncensored models and community fine-tunes, chosen and hosted by independent providers, on an OpenAI- and Anthropic-compatible endpoint. Keep your SDK, change the base URL.

For everyone

Private prompts, by design

Prompts and outputs are decrypted in-process and never logged, written to disk, or stored. Apple hardware attestation and an APNs app-targeted nonce challenge are live for admitted providers; signed per-response receipts remain in progress.

An experimental alpha, built in the open

Umbra is a small experimental alpha. Every hosted model is a public, open-weight model an independent provider pulls with their own Hugging Face key, run in-process behind a confidential coordinator that never sees prompt text. We ship honestly and say what is not ready rather than overstate it.

Read the receipts. Follow the changelog and blog, and see how privacy is enforced in the trust model, attestation, and threat model docs.
Want to try it? Read the quickstart and start free, or bring your idle Mac and become a provider. Deep dive: threat model.