Questions, answered plainly.

Because plans never bill for inference. A ZeroFare subscription entitles you to cloud services — the console, hosted cache, support tiers, SSO. Model spend is metered separately and billed at provider rates. The plan system is implemented so a plan literally cannot include spend credit: the type is grantsProviderSpend: false.

ZeroFare makes money on the service layer, not by marking up your tokens.

Management console access and edge distribution. You can run the entire stack on your own hardware at $0/mo forever, and self-hosting is Apache-2.0 — no feature ransom.

Same API, two deployment shapes. The edge runs on your machine (loopback 127.0.0.1:8787), vaults provider keys locally, and never sends them anywhere. The hosted gateway is ours to operate: you authenticate with a ZeroFare API key over TLS and don't manage provider keys at all. Your code speaks the same contract to either.

Yes — the edge and gateway serve /v1/chat/completions, /v1/responses, and /v1/messages under the versioned zerofare.compat/v1 profile. Errors are structured (code, message, retryable, profile) so clients can react programmatically.

Every request that passes admission produces a decision record — admission outcome, routing, finalization. You can fetch the receipt for any decision your key made, via zf.receipt(decisionId) in the TypeScript SDK or the REST surface. It's the audit trail for "why did my request behave this way?"

Nowhere as plaintext. Gateway API keys are stored as peppered scrypt hashes; management credentials as salted SHA-256 digests. The plaintext is shown to you exactly once at creation and never written to disk — not even in the seed file the services hydrate from.

Yes. The edge, gateway, and console all ship as hardened containers with a compose file targeting Oracle Cloud's Always Free tier (works on any VM with Docker). See Self-hosting for the full procedure.

The hosted path is running today with real authentication, admission, and receipts, but full commercial provider dispatch and metered billing are Phase 2 (in progress). We label every surface honestly — see the roadmap for exactly what's shipped vs. planned.

We don't — not in place. Contracts are versioned (zerofare.compat/v1, zerofare.management/v1) and snapshotted in the repo; a breaking change ships as a new version. Your integration keeps speaking v1.

The business model. Router companies profit per token, so their incentives point toward more expensive routes. ZeroFare's revenue is the flat service subscription, so the margin on your spend is structurally zero — enforced in code, not just policy.