azure-openai-proxy

A tiny Bun proxy that lets OpenAI SDK-based tools (like OpenCode) talk to Azure OpenAI endpoints — including Responses API models (Codex, etc.) that the SDK can't reach directly.

Why?

Azure OpenAI requires ?api-version= on every request. The OpenAI SDK doesn't add this — especially for the Responses API (/responses endpoint used by Codex models). This proxy sits in between and fixes the URL on the fly.

API	SDK sends	Azure expects
Chat Completions	/deployments/{model}/chat/completions	Same + ?api-version=
Responses	/responses (model in body)	/openai/responses?api-version=

Quick start

git clone https://github.com/grumpyowlbear/azure-openai-proxy.git
cd azure-openai-proxy

export AZURE_OPENAI_BASE_URL="https://YOUR-RESOURCE.cognitiveservices.azure.com/openai"
bun proxy.mjs

The proxy listens on http://localhost:18924 by default.

Configuration

Variable	Default	Description
AZURE_OPENAI_BASE_URL	(required)	Your Azure OpenAI base URL
AZURE_API_VERSION	2025-03-01-preview	Azure API version string
UPSTREAM_TIMEOUT_MS	1800000	Hard cap on the outbound fetch to Azure in milliseconds (30 min default, 0 disables)
PORT	18924	Local port to listen on

OpenCode setup

Copy example-opencode-config.json to ~/.config/opencode/config.json and fill in your values.

Use two providers — one direct to Azure for Chat Completions, one through the proxy for Responses API models:

{
  "provider": {
    "azure-cognitive-services": {
      "options": {
        "apiKey": "YOUR_KEY",
        "baseURL": "https://YOUR-RESOURCE.cognitiveservices.azure.com/openai",
        "useCompletionUrls": true,
        "useDeploymentBasedUrls": true
      },
      "models": {
        "gpt-5.3-chat": { "name": "GPT 5.3 Chat" }
      }
    },
    "openai": {
      "options": {
        "apiKey": "YOUR_KEY",
        "baseURL": "http://localhost:18924"
      },
      "models": {
        "gpt-5.1-codex-mini": { "name": "GPT 5.1 Codex Mini" }
      }
    }
  }
}

Switch models in OpenCode with /model.

How it works

OpenCode → openai provider → http://localhost:18924/responses
                                        ↓
                              proxy appends ?api-version=
                              proxy converts Bearer → api-key header
                              proxy streams SSE responses through
                                        ↓
                              https://YOUR-RESOURCE.../openai/responses?api-version=...
                                        ↓
                                   Azure OpenAI ✓

Known issues

• developer role bug: @ai-sdk/openai assumes certain model IDs are "reasoning models" and sends role: "developer" instead of role: "system". This can break non-OpenAI models on Azure when using the Responses API path. Chat Completions path (useCompletionUrls: true) is not affected.

License

MIT