Request metadata

Send the X-Gateway-Metadata: enabled header and the response gains one extra top-level block, gateway_metadata. Without the header the block is absent and nothing else changes, so it is safe to enable per request, for part of your traffic, or only while debugging. The block is also returned under its former name, omnirelay_metadata, with the same content, and the former header name X-OmniRelay-Metadata stays accepted (see Compatibility).

curl https://api.omnirelay.ai/v1/chat/completions \
  -H "Authorization: Bearer $OMNIRELAY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Gateway-Metadata: enabled" \
  -d '{"model": "gpt-6-luna", "messages": [{"role": "user", "content": "Hello"}], "tags": {"client": "acme"}}'

The value is case-insensitive. Any value other than enabled is treated as off.

What you get

{
  "id": "chatcmpl-6f1c…",
  "object": "chat.completion",
  "choices": [{"...": "..."}],
  "usage": {"...": "..."},
  "gateway_metadata": {
    "request_id": "6f1c…",
    "requested": "gpt-6-luna",
    "model": "gpt-6-luna",
    "provider": "OpenAI",
    "mode": "live",
    "strategy": "direct",
    "attempt": 1,
    "attempts": [{"provider": "OpenAI", "model": "gpt-6-luna", "mode": "live"}],
    "tags": {"client": "acme"}
  }
}

Field	Meaning
`request_id`	The request ID, the same as the `X-Request-ID` response header. Quote it to support
`requested`	The model ID you sent
`model`, `provider`	The model that served the call and its provider
`mode`	`live`: a real provider was called
`strategy`	How the model was resolved. Each model has one channel today, so this is always `direct`
`attempt`, `attempts`	Which attempt served and every attempt made. There is no fallback yet, so there is exactly one
`tags`	The Tags recorded for this call after merging the key’s defaults, the header and the body, or `{}`
`deprecated_header`	Only when the request sent `X-OmniRelay-Tags` or `X-OmniRelay-Metadata` together with its current name carrying a different value: the old header that was ignored

strategy, attempt and attempts are there so the shape does not change when routing and fallback arrive. They report what actually happened, never a plan.

On a stream

With stream: true the block rides the first chunk, because the model and provider are settled before any token is generated. Later chunks do not repeat it.

In the console

The API debug tool on the Playground page always requests metadata, so the block appears in its response panel without the header.

V1 Documentation

Quick Start

User manual

Overview

LLMs

Images

Videos

General

What you get

On a stream

In the console

​What you get

​On a stream

​In the console

What you get

On a stream

In the console