An OpenAI-compatible API: point base_url at us and one key reaches 117 models across two ecosystems.
Each call writes a ledger row as it happens — model, tokens, cost, latency. Auditable, billable, allocatable.
Whatever the vendor lists, that is what you pay — no markup, no discount, no multiplier; when a vendor changes its price, we follow. So we are not "the cheaper route": on the same model, our price is the same number you would pay buying from the vendor directly. Token pass-through is not our business — the gateway is the delivery layer for our transformation work, which is our only revenue. Full pricing details live on the Bootcamp page →
One key, 117 models. Switching is a one-field change. No vendor lock-in.
Model, tokens, cost, latency, caller, policy hits — every call, entered in the ledger. The audit is the product, not a log file.
Permissions, quotas and content policy are enforced at the gateway — governance never relies on the model behaving.
Split billing by team and project. Budget caps, overage alerts, metering down to the token.
A request passes, in order: key validity → expiry → source IP → rate limit → allowed models → balance and budget → policy. A call stopped at a gate still writes a ledger entry naming the gate that stopped it — so you debug it yourself instead of writing to us.
Set per key: expiry, source IP allowlist (IPv4/CIDR), per-minute limit, allowed models. A refusal on IP echoes back the source IP of that request, so you can paste it into the allowlist and unblock yourself. Only admins mint keys or change these.
Model allowlist, blocked terms, data egress (email, phone, national id, bank card, API key) and rate limit. All four are evaluated in the gateway; a hit is refused and the ledger records which policy did it. Rate-limit hits return 429 + retry-after, the rest 403.
Every entry carries an initiator and a workflow, both filterable in the ledger and the audit view. Bind a key to a workflow and your client code does not change at all. Machine keys have no person behind them, so we leave that blank rather than invent one.
A model name hard-coded in your code can be mapped to another model with zero client changes. When a mapping applies, the ledger records both the name you asked for and the model it actually ran on — we never swap a model without leaving a trace.
No shelf page. Once connected, GET /v1/models returns the full model list live, across two model ecosystems (international catalog as an authorized AWS reseller). What's available today — and tomorrow — is whatever the endpoint returns.
Token-metered and live: /v1/chat/completions /v1/responses /v1/embeddings /v1/rerank — every call lands on your ledger.
Image generation and speech (billed per image / per character / per second) are off by default: our price table is token-only today, and we would rather disable them than serve them at zero cost and hand you a ledger that never charges. Ask us to enable them — calls are still recorded, flagged "unpriced".
Every delivery engagement's ledger starts from this layer's audit chain. The gateway runs in production on real traffic — that is why we can afford to charge on results.