OpenAI Agents SDK

The base URL goes on an AsyncOpenAI client, parameter base_url, which is then handed to the SDK. Three scopes take it: set_default_openai_client globally, a ModelProvider through RunConfig per run, or Agent.model per agent.

pip install openai-agents

from agents import Agent, Runner, AsyncOpenAI, OpenAIChatCompletionsModel, set_tracing_disabled

set_tracing_disabled(disabled=True)

client = AsyncOpenAI(api_key="sk-...", base_url="https://api.hanzo.ai/v1")
model = OpenAIChatCompletionsModel(model="zen5", openai_client=client)

agent = Agent(name="Assistant", instructions="Answer precisely.", model=model)
result = Runner.run_sync(agent, "Hello")
print(result.final_output)

Lands on POST /v1/chat/completions.

The SDK's own default is the Responses API, so a plain agent would reach POST /v1/responses instead. Both are served, so either works — the snippet uses OpenAIChatCompletionsModel to make the choice visible. The other documented way to make it is set_default_openai_api("chat_completions"), which applies when the address and key come from OPENAI_BASE_URL and OPENAI_API_KEY.

Tracing uploads to OpenAI. With no platform.openai.com key in the environment, disable it as above or every run logs a tracing failure.

MultiProvider(openai_base_url=..., openai_api_key=..., openai_prefix_mode="model_id", unknown_prefix_mode="model_id") is the route for prefix-based dispatch where the endpoint expects literal namespaced ids.

MCP

from agents import Agent
from agents.mcp import MCPServerStreamableHttp

server = MCPServerStreamableHttp(params={
    "url": "https://api.hanzo.ai/v1/mcp",
    "headers": {"Authorization": "Bearer sk-..."},
    "timeout": 10,
})

agent = Agent(name="Assistant", instructions="Answer precisely.", model=model, mcp_servers=[server])

There is also HostedMCPTool, which passes a server_url in tool_config — that round trip is run by the OpenAI Responses API, not by your process.