Reference
Gateway endpoints and headers
The routes the gateway serves, the headers it reads and the headers it sends back.
Base URL: https://ai.odux.io/gateway/v1, or https://ai.odux.io/gateway/agent/<name>/v1 to name the agent in
the address. For Anthropic clients, which add /v1 themselves, leave it off: https://ai.odux.io/gateway/agent/<name>.
#Routes
| Route | Format | For |
|---|---|---|
POST /v1/chat/completions |
OpenAI Chat Completions, streamed or not | Most agents and SDKs |
POST /v1/responses |
OpenAI Responses | Codex |
POST /v1/messages |
Anthropic Messages | Claude Code, the Anthropic SDK |
POST /v1/messages/count_tokens |
Anthropic token counting | Claude Code |
GET /v1/models |
OpenAI model list | Model pickers: only the models your stored keys allow |
GET /v1/whoami |
JSON | Who requests with this credential are filed under, and what to fix if no one |
GET /health |
JSON | Whether the gateway is up (no key) |
A provider takes a format only when it has that endpoint: Anthropic Messages goes to Anthropic, and to DeepSeek and
Kimi through their Anthropic-compatible endpoints; Responses goes to OpenAI. Any other model gets a clear 400 in the
format's own error shape.
#Headers it reads
| Header | Meaning |
|---|---|
x-api-key or Authorization: Bearer |
Your IA4AI key. Both are accepted. |
x-api-key with a provider key as Authorization: Bearer |
Your IA4AI key, plus your own provider key to forward without storing it. |
x-ia-key |
Your IA4AI key, for clients whose API-key field must hold the provider key (such as ANTHROPIC_CUSTOM_HEADERS in Claude Code). |
x-ia-agent |
The agent's name on your dashboard. A label, not a credential. |
#Headers it sends back
| Header | Meaning |
|---|---|
x-ia-attribution |
client when the call was filed under your workspace, unattributed when it was not |
x-ia-workspace |
The workspace the call was filed under |
x-ia-attribution-reason |
Why a call was not attributed |
x-ia-agent |
The agent name the gateway gave the call |
x-ia-recommended-model, x-ia-route-band |
When the call looks like one a smaller model could handle: the suggestion. Advisory only. |
#What is exported to the engine
For each call, an OpenTelemetry GenAI span with the model, provider, token counts (input as the whole prompt, cache reads and writes counted), latency, tool names and the agent. Never the prompt, the answer or tool arguments, unless your workspace's admin has content mode on for that kind.

