PRODUCT · SOVEREIGNROUTER
A drop-in replacement for the OpenAI API endpoint that adds tier-gated routing, BYOK tenancy, automatic provider fallback, and per-call attestation written to your audit log. Every other change to your code is zero.
drop-in replacement
Existing OpenAI client code, unmodified except for one URL. The router speaks the OpenAI chat-completions schema bit-for-bit, plus an X-SF-Policy header for the trust tier and an _sf_attestation field on the response.
how it works
Your sk-router-... key resolves to a tenant with allowed policies, spend cap, and rate limits. Multi-tenant by design; one router instance serves the whole org.
Inbound prompt scanned for OpenAI / Anthropic / AWS / GitHub / Slack / Stripe key formats. Findings logged to rekor; redacted in-place before the prompt reaches any provider.
Policy specifies required tier, required witness scopes, allowed origins, drift posture, latency cap, cost cap, and preferred provider chain. Models that don't qualify are rejected before any cost is incurred.
Eligible providers ordered by health score (rolling p50/p95 + success rate). First failure walks to the next; second failure walks again. Audit row records every attempt.
Real adapter for Anthropic, OpenAI, Bedrock (with SigV4), Vertex, Ollama, vLLM. Your BYOK provider keys are read from the keystore chain (LocalFile → Env → KMS) and never appear in logs.
Response augmented with _sf_attestation containing model hash, provider ID, policy ID, route decision ID, rekor index, mirror ID, and tier observed. Audit row appended to the Merkle log atomically.
at the numbers
sf_attestation event lands at the end of the stream alongside [DONE].how it differs
Other gateways route by latency, price, or capability. SovereignRouter routes by tier. clinical-strict won't pick a non-FDA-witnessed model even if it's cheaper. Policy is a first-class object, not a config flag.
Every response carries a rekor index. Your audit pipeline can prove what model produced what output, what witnesses signed it, what tier it was at the moment of the call. Other gateways don't even know.
Provider keys never appear in routing decisions, telemetry, logs, or fallback chain inspection. Keystore chain (LocalFile → Env → KMS) keeps them out of the dataplane entirely. Tenant keys are short-lived and rotatable.
pricing
You pay providers directly through your BYOK keys. We charge a margin on the inference passing through. Self-hosted, you pay nothing — the binary and Helm chart are yours.
Run the binary in your own VPC. Full feature set. Source available, Apache 2.0.
We host. We monitor. You point your clients at our endpoint and bring your own provider keys.
Dedicated tenancy, private network connectivity, bespoke witness panel integration.
Run it yourself with the open-source binary. Or let us host it. Either way, the trust substrate is yours.