Umbra is an experimental alpha, and the fastest way to lose trust in a private inference product is to overclaim. So here is a plain snapshot of what actually works today, and what does not yet.
What works today
You can call the API. Umbra speaks the OpenAI and Anthropic wire formats.
Point your existing client at https://api.tryumbra.dev/v1, mint a key in the
console, call GET /v1/models to see what is live, and run it. Your $1 of free
signup credit covers up to ~10 million tokens at the current lowest input
rate ($0.10 per million). The exact count depends on the model and your
input/output mix. See pricing for the live per-model rates.
Developer top-up is live. You can buy prepaid inference credit with a card, processed by Stripe. Charges are real. Purchased and promotional credit is spendable on inference only. It is non-refundable and non-withdrawable.
Attestation is live at the hardware tier. The coordinator runs inside an
attested AMD SEV-SNP confidential VM, and Apple Managed Device Attestation of
provider hardware is live: the reference provider runs at the hardware trust
tier under real MDA verification. Prompts are decrypted in memory only and
zeroized after each request, never logged, never written to disk.
The provider app is published. A clean Apple-Silicon Mac can review the publisher identity and every system change on the download page, install the signed and notarized DMG, complete the separately approved MDM enrollment, and start hosting a public Hugging Face GGUF model it selects and pulls with its own Hugging Face key.
Provider payouts are rolling out. Connect from the provider wallet when it shows payouts available, complete Stripe-hosted identity and bank verification, and withdraw settled earnings once you clear the minimum. Projections are not payable. Region and eligibility details live in the provider agreement.
What is still pending
Independent code identity (code_attested) is live. After MDA verifies the
hardware tier, an APNs app-targeted nonce challenge proves the genuine entitled
provider app can receive and answer it. Authenticated private routing requires
that tier and fails closed below it rather than silently downgrading.
Signed per-response receipts are not live yet. Today’s evidence covers coordinator posture and provider registration. The receipts that would bind a specific request to that evidence are still rolling out. Treat missing evidence as unavailable privacy proof.
Try it
Swap your base_url, keep your SDK: quickstart. See the live
rates on pricing, or read the trust model and
threat model for how the privacy property is meant to hold.
(Umbra is an experimental alpha. Everything above reflects the state at the time of writing and will change as the alpha progresses: see the changelog.)