AutoGen

OpenAIChatCompletionClient takes base_url, and model_info alongside it whenever the model name is not one OpenAI serves.

OpenAIChatCompletionClient takes base_url, documented as required when the model is not hosted on OpenAI. The part that catches people: any model name the SDK does not recognise also needs model_info, and without it the client raises rather than calling.

pip install "autogen-ext[openai]"

from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_core.models import ModelFamily

client = OpenAIChatCompletionClient(
    model="zen5",
    base_url="https://api.hanzo.ai/v1",
    api_key="sk-...",
    model_info={
        "vision": False,
        "function_calling": True,
        "json_output": True,
        "family": ModelFamily.UNKNOWN,
        "structured_output": True,
    },
)

# ... use the client, then release it
await client.close()

Lands on POST /v1/chat/completions.

Where a model's record in GET /v1/models declares supports_vision or supports_tools, those map to vision and function_calling. Most records declare neither, and AutoGen's reference states plainly that this client is not tested against non-OpenAI models — so claim only what you have exercised.

MCP

The URL is the url field of StreamableHttpServerParams.

from autogen_ext.tools.mcp import StreamableHttpMcpToolAdapter, StreamableHttpServerParams

params = StreamableHttpServerParams(
    url="https://api.hanzo.ai/v1/mcp",
    headers={"Authorization": "Bearer sk-..."},
    timeout=30.0,
    sse_read_timeout=300.0,
    terminate_on_close=True,
)

adapter = await StreamableHttpMcpToolAdapter.from_server_params(params, "describe")

StdioServerParams and SseServerParams are the siblings; mcp_server_tools() builds every tool on a server in one call rather than one adapter at a time.

MCP

On this page