Drop-in OpenAI-compatible endpoint. Strip waste, redact PII, route to the cheapest capable model, cap spend before the bill, return the savings math on every call.
PII → strip → prune → optional semantic compress. Free steps first. Paid rewrite only if still over target.
cheapest / balanced / quality. Transparent fallback. No silent downgrade.
Per-key daily and monthly budgets, reserved before the upstream call. Over cap is HTTP 402.
Three-baseline savings: naive, self-reported, same-model. CSV for the CFO.