OpenAI

Configuration

credentials:
  - name: "openai_main"
    type: "openai"
    api_key: "sk-proj-xxxxx"
    base_url: "https://api.openai.com"
    rpm: 100
    tpm: 50000

Required Fields

Azure OpenAI

For Azure OpenAI, use the same openai type with the Azure endpoint:

credentials:
  - name: "azure_openai"
    type: "openai"
    api_key: "os.environ/AZURE_OPENAI_KEY"
    base_url: "https://your-resource.openai.azure.com/openai/v1"
    rpm: 100
    tpm: 50000

For native Responses WebSockets, configure the deployment name and enable the transport:

models:
  - name: gpt-6-astra
    model: your-astra-deployment
    credential: azure_openai
    passthrough_responses: true
    websocket_responses: true

The router connects to wss://your-resource.openai.azure.com/openai/v1/responses with the credential's API key as a Bearer token. See the Azure WebSocket guide and router session limits.

Per-Model Configuration

Use the models section to control per-model behavior:

models:
  - model: "gpt-4o-mini"
    rpm: 500
    tpm: 200000

  - model: "gpt-5.3-codex"
    passthrough_responses: true   # forward Responses API requests natively (auto-detected for codex models)

  - model: "gpt-4o"
    passthrough_responses: false  # force Chat Completions conversion even if model name contains "codex"

passthrough_responses

Controls whether Responses API (/v1/responses) requests for this model are forwarded natively to the provider's /v1/responses endpoint instead of being converted to Chat Completions format.

This is useful when deploying non-codex models that natively support /v1/responses, or to force conversion for codex-named models behind a Chat Completions-only endpoint.

Responses API

Requests to /v1/responses are automatically detected and handled:

• Standard models (gpt-4o, gpt-4o-mini, o1, o3, etc.): converted to Chat Completions format internally, response converted back. Works with Azure OpenAI and all other OpenAI-compatible backends.
• Codex / passthrough models: forwarded natively to the provider's /v1/responses endpoint without conversion.

See Responses API documentation for full details.

Web Search Tools

Tool handling follows the outbound endpoint contract:

• Native /v1/responses passthrough keeps built-in tools and tool_choice unchanged, including web_search, file_search, computer_use, and code_interpreter.
• /v1/chat/completions converts web_search and web_search_preview to top-level web_search_options for models whose name contains search-preview, including search_context_size and user_location.
• Chat Completions requests for other models preserve web_search, web_search_preview, and matching tool_choice values for upstream validation. Ordinary function tools in the same request also remain unchanged.

# Converted to web_search_options for Chat Completions
client.chat.completions.create(
    model="gpt-4o-search-preview",
    tools=[{"type": "web_search_preview", "search_context_size": "high"}],
    ...,
)

# Built-in tool is preserved for upstream validation
client.chat.completions.create(model="gpt-5.6-sol", tools=[{"type": "web_search"}], ...)

Other non-function tool types (e.g. computer_use, code_execution) are also dropped for OpenAI Chat Completions requests. They are handled by their respective providers (Vertex AI, Anthropic) when those backends are configured.