Relay gateway

The Relay gateway receives the model request, applies the project’s policy, calls the selected provider with a connected provider credential, and returns the provider-compatible response.

Send a request

Call Relay’s OpenAI-compatible endpoint directly:

curl https://relay.oximy.com/v1/chat/completions \
  -H "Authorization: Bearer ox_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "oximy/auto",
    "messages": [{"role": "user", "content": "Summarize the release risk."}]
  }'

POST /v1/chat/completions HTTP/1.1
Host: relay.oximy.com
Authorization: Bearer ox_...
Content-Type: application/json

{
  "model": "oximy/auto",
  "messages": [
    {"role": "user", "content": "Summarize the release risk."}
  ]
}

Use the gateway URL shown in your Relay setup when it differs from the public URL above. Because the endpoint is OpenAI-compatible, an existing compatible client can use it by replacing its base URL and API key. Relay does not require an Oximy client library.

Select a model

Send oximy/auto to apply Relay’s active routing policy.
Send a provider-prefixed model ID, such as openai/gpt-4o-mini, to request that model directly.
Configure a fallback model for automatic-routing failures.

Provider-prefixed IDs avoid ambiguity where multiple providers expose similar model names.

Production checklist

Use a dedicated project and key per application environment.
Verify each provider credential and confirm its models are callable.
Set project and key limits before increasing traffic.
Exercise the fallback with a controlled failure.
Confirm Traffic records show the expected source, resolved model, and routing reason.
Revoke old keys after rotation.

A Relay key authenticates the application to Relay. It is not a provider API key, and it should never be sent directly to a model provider.

Configure routing and fallbacks

Define the candidates and behavior behind oximy/auto.

HTTP endpoint reference Observe and OTEL

⌘I

​Send a request

​Select a model

​Production checklist

Configure routing and fallbacks

Send a request

Select a model

Production checklist