Agno

OpenAILike takes base_url. A sibling class reaches the responses shape on the same parameter.

OpenAILike takes base_url. It extends the OpenAI-compatible interface, so every parameter of OpenAIChat works on it too.

pip install agno openai

from agno.agent import Agent
from agno.models.openai.like import OpenAILike

model = OpenAILike(
    id="zen5",
    api_key="sk-...",
    base_url="https://api.hanzo.ai/v1",
)

agent = Agent(model=model)
agent.print_response("Share a 2 sentence horror story.")

Lands on POST /v1/chat/completions.

id defaults to not-provided, so it is the one field you must set. OpenResponses from agno.models.openai is the sibling on the same base_url parameter and reaches POST /v1/responses.

MCP

MCPTools takes transport and url.

from agno.agent import Agent
from agno.tools.mcp import MCPTools

mcp_tools = MCPTools(
    transport="streamable-http",
    url="https://api.hanzo.ai/v1/mcp",
    headers={"Authorization": "Bearer sk-..."},
)

agent = Agent(id="hanzo-agent", name="Hanzo Agent", model=model, tools=[mcp_tools])

Two things the docs flag. Do not serve AgentOS with reload=True while MCPTools is attached. And the connection is not refreshed on its own — set refresh_connection=True for a long-lived agent.

MCP

On this page