Connecting

Integration

Connect any OpenAI-compatible client tohttps://your-gateway.example.com/v1

Connection

Use these values in any OpenAI-compatible client.

Base URL

—

Model

deepseek-v4.1-flash

Authentication

Authorization: Bearer <gateway key>

Protocol

OpenAI Chat Completions · SSE streaming

Payload cap

—

Models

Requests for any model id are served by the active model; this is the canonical id.

Loading model metadata…

Examples

Copy-paste snippets for common OpenAI-compatible clients.

POST /v1/chat/completions

curl https://your-gateway.example.com/v1/chat/completions \
  -H "Authorization: Bearer $GATEWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'

Endpoints

Public routes served by the gateway.

Method	Path	Description
POST	/v1/chat/completions	Chat completions (streaming, tools, vision, reasoning)
GET	/v1/models	OpenAI-compatible model list
GET	/health	Liveness probe
GET	/status	Queue and upstream status JSON

OpenAI compatibility

What is translated and preserved.

Chat completions. system/developer, user, assistant and tool messages are normalized and forwarded.
Streaming. SSE with OpenAI chat.completion.chunk events, usage and data: [DONE].
Vision. image_url parts as data URLs, base64 or https URLs are translated to the provider's image blocks.
Tool calling. OpenAI tools/tool_choice are translated; tool_calls and tool results round-trip.
Reasoning. reasoning_effort or thinking enables reasoning; reasoning_content is returned in messages and deltas.
Usage. Input, output and cached prompt tokens are reported, including on streamed responses.
Errors. Standard OpenAI error objects with provider status codes preserved.

Operational notes

Behaviour worth knowing when integrating.

The gateway keeps a strict FIFO queue and holds one upstream request at a time, so concurrent clients never see provider concurrency errors. Excess latency is queue wait, visible per request.

Provider fair-use 429 responses pause the lane for the advertised Retry-After; queued requests resume automatically and the affected request is retried.

n > 1 is not supported and returns a clear error. Sampling penalties and logprobs are accepted but not forwarded. Remote image URLs and base64/data-URL images are both supported.

The upstream provider key is held server-side only; clients authenticate exclusively with the gateway bearer token.