CrewAI

LLM takes base_url. custom_openai=True is what lets you name a model id CrewAI does not recognise.

LLM takes base_url. Prefix the model with openai/ so LiteLLM routes it through the OpenAI-compatible path.

pip install crewai

from crewai import LLM, Agent

llm = LLM(
    model="openai/zen5",
    base_url="https://api.hanzo.ai/v1",
    api_key="sk-...",
)

agent = Agent(
    role="Researcher",
    goal="Answer precisely",
    backstory="You check before you answer.",
    llm=llm,
)

Lands on POST /v1/chat/completions.

For a model id CrewAI does not recognise as an OpenAI name, add custom_openai=True — that flag is what lets a gateway model id through unrewritten:

llm = LLM(
    model="anthropic/claude-sonnet-4-6",
    custom_openai=True,
    base_url="https://api.hanzo.ai/v1",
    api_key="sk-...",
)

The environment equivalents are OPENAI_API_KEY and OPENAI_BASE_URL.

MCP

MCP servers are first-class on the Agent, in an mcps list. The URL is the url parameter.

from crewai import Agent
from crewai.mcp import MCPServerHTTP

server = MCPServerHTTP(
    url="https://api.hanzo.ai/v1/mcp",
    headers={"Authorization": "Bearer sk-..."},
    streamable=True,
    cache_tools_list=True,
)

agent = Agent(role="Researcher", goal="Answer precisely", backstory="", llm=llm, mcps=[server])

For manual lifecycle control there is MCPServerAdapter from crewai_tools, installed with pip install 'crewai-tools[mcp]', which takes a dict of url and transport.

MCP

On this page