OpenAI

VM-X uses the official openai SDK to talk to api.openai.com.

Connection config

Field	Required	Description
`apiKey`	yes	Standard OpenAI API key. Create one at platform.openai.com → API keys.

{
  "provider": "openai",
  "config": { "apiKey": "sk-..." }
}

Endpoint passthrough

OpenAI is the only provider that natively speaks both Chat Completions and Responses, so two of the three VM-X endpoints forward verbatim. Anthropic Messages takes a single direct hop through the Responses converter (no internal pivot):

Client request shape	What hits the wire
Chat Completions	Verbatim. Body is forwarded to `client.chat.completions.create()` with the gateway envelopes (`vmx`, `__vmx_passthrough`) stripped.
Responses	Verbatim. Body is forwarded to `client.responses.create()` after stripping envelopes. Native `Response` / `ResponseStreamEvent` shape on the way back.
Anthropic Messages	Direct converter to OpenAI Responses (not Chat Completions — Responses' richer event vocabulary preserves `thinking` / hosted tools / structured output).

Capabilities

Capability	Status
Streaming	✅
Function / tool calling	✅
Structured outputs (`json_schema`)	✅
Vision (`image_url` / `input_image`)	✅
Reasoning models (o-series)	✅ — `reasoning_tokens` flow through audit/usage
Web search	✅ — search-class models (`gpt-4o-search-preview`, `gpt-4o-mini-search-preview`) emit `annotations` with `url_citation` blocks
Predicted outputs	✅
Audio I/O	✅

`providerArgs` — common OpenAI-native fields

{
  "vmx": {
    "providerArgs": {
      "service_tier": "scale",
      "user": "user-7e9...",
      "logit_bias": { "1234": -100 }
    }
  }
}

Models

VM-X doesn't restrict the model list — pass any model id OpenAI accepts (gpt-4.1, gpt-4o, gpt-4o-mini, o3-mini, o4-mini, gpt-4o-search-preview, …). The model is forwarded to the SDK verbatim. The default for new connections is gpt-4.1.

Notes

Per-model maxRetries is forwarded to the OpenAI SDK as per-call maxRetries, so transient 5xx / throttling errors retry inside the SDK before the gateway falls through to the next fallback model. See AI Resources.
Rate-limit headers (x-ratelimit-*) are parsed and surfaced as retry-after on 429 CompletionErrors.
vmx envelope strip: OpenAI rejects unknown top-level fields with a 400. The gateway strips both vmx (correlationId / metadata / providerArgs) and __vmx_passthrough (cross-format carrier) before send, on every endpoint.

Connection config​

Endpoint passthrough​

Capabilities​

providerArgs — common OpenAI-native fields​

Models​

Notes​