openai-responses-adapter

A lightweight Go proxy that allows clients to call the OpenAI Responses API style endpoint, while the adapter forwards requests to an upstream Chat Completions API provider.

Current status

This is an MVP scaffold with:

• POST /v1/responses (Responses -> Chat Completions mapping)
• GET /v1/responses/{response_id} (read stored response by id)
• GET /healthz
• previous_response_id multi-turn context chaining
• function calling mapping (tools, tool_choice, function_call_output subset)
• MCP tool bridge:
  • load MCP servers from config file
  • aggregate MCP tools as internal native tools
  • rewrite native tools to upstream function tools
  • auto-execute MCP tools when upstream emits tool calls (non-stream requests)
• stream=true SSE passthrough with Responses-style delta events
• in-memory response store with TTL + LRU
• passthrough auth header or fallback to UPSTREAM_API_KEY

Current MVP request constraint:

• model is required
• input supports: string or item array (message / function_call_output subset)
• downgrade_developer_to_user is optional:
  • default true: convert outbound role=developer to role=user for upstream compatibility
  • set false to keep developer role unchanged
• tool compatibility:
  • function is supported directly
  • non-function/native tool types are supported when they match configured MCP tool names
• streaming currently covers: response.created, response.in_progress, response.output_item.added, response.output_text.delta/done, response.function_call_arguments.delta/done, response.output_item.done, response.completed, [DONE] (MVP subset)

Run

go run ./cmd/adapter

Build Docker image (local)

docker build -t openai-responses-adapter:dev .

Environment variables

• LISTEN_ADDR (default :8080)
• UPSTREAM_BASE_URL (default https://api.openai.com)
• UPSTREAM_CHAT_PATH (default /v1/chat/completions)
• UPSTREAM_API_KEY (optional fallback if request has no Authorization)
• STORE_MAX_ENTRIES (default 1000)
• STORE_TTL (default 1h, Go duration format such as 30m, 2h)
• MCP_CONFIG_PATH (optional path to MCP server config JSON)

MCP config example

Claude mcpServers format:

{
  "mcpServers": {
    "web-search-prime": {
      "type": "http",
      "url": "https://open.bigmodel.cn/api/mcp/web_search_prime/mcp",
      "headers": {
        "Authorization": "Bearer your_api_key"
      }
    }
  }
}

Release artifacts

When a GitHub Release is published, GitHub Actions will run GoReleaser and publish:

• Multi-platform binaries to the release assets:
  • linux/darwin/windows
  • amd64/arm64
• Docker images to GHCR:
  • ghcr.io/kurok1/openai-responses-adapter:<tag>

Example

curl -sS http://localhost:8080/v1/responses \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer <token>' \
  -d '{
    "model":"gpt-4o-mini",
    "input":"你好，介绍一下你自己",
    "downgrade_developer_to_user": true,
    "stream":false
  }'