Use case · On-prem & air-gapped

Frontier-class AI, air-gapped.

Self-host the gateway and open models where nothing leaves the wire — the same control bands, fully disconnected.

The problem

The best models live in someone else’s cloud.

For classified, OT, or sovereignty-bound environments, calling a hosted model API is a non-starter. But the mission still needs capable AI — and bolting a raw open-source model onto the network with no gateway, no guardrails, and no log is its own kind of exposure.

// The pattern

The whole control plane, disconnected.

The same gateway, guardrails, egress control, and audit trail run entirely inside your enclave against self-hosted open models. Default-deny egress isn’t a setting you can forget — it’s enforced by the network and the proxy. Nothing phones home.

// 01 · No egress, by construction

The environment can’t reach out.

Security-group rules and an allowlist proxy mean the enclave can’t touch the open internet even if something tries to.

// 02 · Local guardrails

Policy runs on your hardware.

Detectors and guardrail policy execute in your enclave, pre-call — no dependency on an external moderation service.

// 03 · Your evidence, your storage

Nothing leaves for a vendor console.

The structured audit log lives in your enclave, on your retention schedule, readable without an outside service.

What stays with you

The boundary, drawn explicitly.

Optimal builds and hands over the pattern. Your enclave, your models, your data, and your authorization decisions remain entirely yours.