Bypass OpenAI Chat Completions

Native protocol and authentication

Use an XHub public model name and virtual key with the matching Bypass deployment. XHub maps the model and injects deployment credentials, preserves native fields and enforces access and billing. Bypass does not run the standard pre-call text guardrail chain. Send Authorization: Bearer $XHUB_API_KEY. Model availability depends on configured deployments.

Protocol behavior

Read choices[].message for normal responses and choices[].delta for streaming. Tool calls require application execution and subsequent result messages. SSE errors after headers cannot change HTTP status. POST /chat/completions is an alias.

Endpoint and authentication

POST /bypass/openai/v1/chat/completions

Use an XHub API key: Authorization: Bearer . Do not use upstream provider credentials as gateway keys.

Use the gateway root Base URL shown in the playground, without appending /v1; send requests to the full endpoint path above.

Request example

curl "$XHUB_BASE_URL/bypass/openai/v1/chat/completions" \
  -H "Authorization: Bearer $XHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"YOUR_MODEL_NAME","messages":[{"role":"user","content":"Hello, introduce yourself."}],"stream":false}'

Request parameters

Field	Type	Description
model	string · required	Public model name; deployment must support this protocol.
messages	object[] · required	role and content; tool results include tool_call_id. Multimodal content requires support.
stream	boolean	False by default; true emits SSE ending with [DONE].
temperature / top_p	number	Sampling; accepted ranges depend on model.
max_tokens / max_completion_tokens	integer	Output ceiling, not spent tokens; field depends on model.
tools / tool_choice	array / string or object	Function definitions and selection; caller executes tools.
response_format	object	Structured output only when deployment supports it.

Nested message and tool fields

Field	Type	Meaning
messages[].role	string	system, developer, user, assistant or tool, subject to model support.
messages[].content	string or object[]	Text or multimodal parts; null assistant content may accompany tool calls.
messages[].content[].type	string	For compatible vision models use text or image_url. Other media types depend on the protocol adapter.
messages[].content[].text	string	Text input.
messages[].content[].image_url.url	string	Public image URL or data URL when supported by the upstream model.
messages[].content[].image_url.detail	string	auto, low or high where supported; not a universal Gemini media-resolution mapping.
messages[].tool_call_id	string	Required for a tool result; matches the assistant tool call ID.
messages[].tool_calls[].id	string	Assistant tool call identifier.
messages[].tool_calls[].function.name	string	Function name.
messages[].tool_calls[].function.arguments	string	JSON-encoded arguments; validate before executing application tools.
tools[].type	string	function.
tools[].function.name	string	Function name supplied by the application.
tools[].function.description	string	Optional function description.
tools[].function.parameters	object	JSON Schema for arguments.
tool_choice	string or object	auto, none, required or a named function where supported.
response_format.type	string	json_object or json_schema where supported.
response_format.json_schema	object	Schema definition, name and strict setting where supported.

Streaming and capability checks

stream defaults to false. With stream=true, consume SSE data events, aggregate choices[].delta.content and tool call argument fragments by index, and stop at [DONE]. Usage may be absent unless the upstream supports it. An error after response headers were sent may appear in an event rather than as a new HTTP status.

Reasoning fields such as reasoning_effort, reasoning and thinking have model-specific formats. Image generation extensions, image_config, safety_settings, media_resolution and file_id mappings are not a universal XHub contract. Native Bypass preserves the corresponding upstream protocol; converted requests must be checked against the implemented adapter. Do not assume the same parameter works across OpenAI, Anthropic, Gemini or other providers.

Success response

{
  "id": "chatcmpl-example",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hello!"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 10,
    "completion_tokens": 5,
    "total_tokens": 15
  }
}

Field	Type
id	string
object	string
choices	object[]
choices[].index	number
choices[].message	object
choices[].message.role	string
choices[].message.content	string
choices[].finish_reason	string
usage	object
usage.prompt_tokens	number
usage.completion_tokens	number
usage.total_tokens	number

Errors and troubleshooting

HTTP	Meaning
400	Invalid request. Check JSON, parameters and query.
401	Missing or invalid XHub credential.
403	Permission denied. Check model scope or policy.
429	Limit exceeded. Inspect response details and Retry-After.