Provider guide

Run models from your Mac.

Umbra providers host public Hugging Face models on Apple Silicon. The console is for operations; this page is the plain setup guide for humans and AI agents. Source-build preview — no prebuilt binary is published yet; you build from source.

Create account Open provider console Read attestation spec

Hardware

Apple Silicon Mac

Memory

36 GB recommended

Model source

Public HF repos

Prompt access

Not logged or persisted

Setup Steps

Source-build preview. No prebuilt binary is published yet. The provider is built from source on your Mac. The curl … | sh below is a checker/guide — it inspects your machine and prints the build steps; it does not install a binary. Read it first.

Check your machine. Run the install checker. It verifies your Mac (Apple Silicon, macOS 15+, memory) and prints the exact build steps — it does not install a binary.
```
curl -fsSL https://tryumbra.dev/install | sh
```
Clone and bootstrap. Clone the repo and run the bootstrap script to fetch the toolchain and dependencies.
```
git clone https://github.com/owenisas/umbra
cd umbra && ./bootstrap.sh
```
Build from source. Build the provider locally with Swift.
```
swift build -c release
```
Pair this Mac. Run the provider login command, then enter the device code in your Umbra account.
```
.build/release/umbra login
```
Add your Hugging Face token locally. Use a read-only token. It is stored on the Mac and used only for public model downloads.
```
.build/release/umbra hf connect
```
Approve a public model. In the provider console, approve a Hugging Face repo and pin its digest before serving it.
Start serving. Routing begins after attestation, memory, and warm-model checks pass.
```
.build/release/umbra serve
```

What Umbra Checks

Before traffic routes

Secure Enclave device key is present.
SIP and Apple attestation checks pass.
The model is warm and its digest matches the approved repo.
Enough memory remains for model weights and KV cache.

What stays local

Your Hugging Face token.
The provider device private key.
Downloaded model weights.
Buyer prompts and outputs after request completion.

AI Agent Notes

Task target: set up a provider Mac. Do not ask for secrets in chat. Ask the operator to paste the Hugging Face token into the local CLI. Never upload the token to Umbra. After setup, use `/console/provider`, `/console/provider-models`, `/console/provider-trust`, and `/console/provider-payouts` for operations.

{
  "provider_setup_url": "/providers",
  "distribution": "source-build preview — no prebuilt binary published yet",
  "source_repo": "https://github.com/owenisas/umbra",
  "console_urls": [
    "/console/provider",
    "/console/provider-models",
    "/console/provider-trust",
    "/console/provider-payouts"
  ],
  "install_checker": "curl -fsSL https://tryumbra.dev/install | sh  (checks the machine + prints build steps; does NOT install a binary)",
  "build_steps": ["git clone https://github.com/owenisas/umbra", "./bootstrap.sh", "swift build -c release"],
  "secret_handling": "HF token stays on the provider Mac"
}