smolagents

OpenAIModel takes api_base, documented as the base URL of the OpenAI-compatible API server. Note the spelling — it is not base_url — and note the class name: the reference documents OpenAIModel, and OpenAIServerModel no longer appears in it.

pip install smolagents

import os
from smolagents import OpenAIModel, CodeAgent

model = OpenAIModel(
    model_id="zen5",
    api_base="https://api.hanzo.ai/v1",
    api_key=os.environ["HANZO_API_KEY"],
)

agent = CodeAgent(tools=[], model=model)
agent.run("How many seconds in a fortnight?")

Lands on POST /v1/chat/completions.

The full signature is OpenAIModel(model_id, api_base=None, api_key=None, organization=None, project=None, client_kwargs=None, custom_role_conversions=None, flatten_messages_as_text=False, **kwargs). Generation arguments — temperature, max_tokens, top_p — are accepted at construction and forwarded on every call. LiteLLMModel is the sibling and takes api_base too.

MCP

MCPClient takes url and transport, and is a context manager that yields the tools.

from smolagents import MCPClient, CodeAgent

with MCPClient({"url": "https://api.hanzo.ai/v1/mcp", "transport": "streamable-http"}) as tools:
    agent = CodeAgent(tools=tools, model=model)
    agent.run("What is my account balance?")

ToolCollection.from_mcp is the other entry point, on the same dict; it requires trust_remote_code=True. streamable-http is the default transport.