Carnegie Relay sits between your applications and Vertex AI, Anthropic, OpenAI or any OpenAI-compatible provider — using your own provider keys. It enforces a model allow-list per environment, hard monthly budgets in dollars, PII and prompt-injection guardrails, and writes a request-level audit trail without storing prompt bodies. Your developers change one base URL. Your IT team gets the controls it was asked for.
One base URL change for developers; policies, budgets and evidence for the people accountable for them.
Anthropic Messages, OpenAI Chat Completions and OpenAI Responses wire formats are accepted unchanged. Set the base URL and a Relay key; no code changes in your application.
Production, staging and each team get their own environment: which models are allowed, which provider keys are used, the monthly budget and whether it is a hard stop or an alert, which guardrails run on requests and responses.
Every request is recorded with model, tokens, latency, cost estimate at the provider's list price, guardrail verdict, environment and key — and no prompt or completion text unless the environment opts in. Nightly export to a bucket or dataset you own.
Each control is per environment, so production and a team's sandbox can have different models, budgets and rules.
Per environment. A request for a model that is not on the list is refused with a clear error naming the list.
Budgets are counted at the provider's list price even on your own keys. Hard mode returns 402 at the limit; soft mode alerts at 80 % and 100 %. The check is concurrency-safe: ten simultaneous requests cannot overspend.
Deterministic PII detection and prompt-injection checks on request and response, per environment policy: flag or block. The verdict is returned in a response header and recorded in the trace.
An ordered channel list per environment — your BYOK keys first, Vertex AI as the Google-resident default. A provider that errors is cooled down for sixty seconds; the next request skips it.
Admin actions and request traces are kept for 30 / 90 / 365 days by plan and exported nightly to your Cloud Storage bucket or BigQuery dataset. Webhook to your SIEM for budget and guardrail events.
Console sign-in with Google Workspace or Microsoft Entra. API keys are scoped to one environment, rate-limited per key, and rotate without downtime.
Annual plans per organisation with included requests; overage per million requests. Model tokens are never on our invoice — they stay with your provider on your own keys.
One platform team, one environment
Company-wide AI platform
Regulated or multi-business-unit groups
Request — One Request = one model call relayed (streaming or not), regardless of tokens. Model tokens are billed by your provider on your own keys; Relay meters requests only.
| Chat completion, 2k tokens | 1 request |
| Streaming completion, 40k tokens | 1 request |
| Request refused by allow-list or budget | 0 requests |
| Guardrail check on request and response | included |
List prices are published when the Marketplace listing goes live. Early-access partners receive their terms in writing before any charge.
The company's industry agents — TieOut Pilot, D&D Pilot, Field Triage and Tender Answer — run on Relay. The same allow-lists, budgets and audit trail apply to them.
By default no text is stored: traces hold metadata (model, tokens, latency, verdicts, cost estimate). An environment can opt in to store bodies for debugging, with its own retention.
So the model charge stays on your provider invoice at your negotiated price, and so Relay has no incentive on token volume. Relay's price is per organisation and per request.
Relay is stateless per request and runs on Cloud Run across zones. The SDK fallback pattern is documented: on a Relay health failure, clients may call the provider directly with their own key — your policy decides whether that is allowed per environment.
Vertex AI (Gemini and Claude on Google Cloud), Anthropic, OpenAI, and any OpenAI-compatible endpoint. Others on request at the Enterprise tier.
On Google Cloud Marketplace as an annual plan per organisation; the charge lands on your Google Cloud invoice and counts toward your commitment. Private offers for Enterprise. AWS Marketplace to follow.
We are onboarding a small number of platform teams. Tell us which providers you use and what controls you are missing; we set up an organisation with your environments.