Google ADK

ADK reaches an OpenAI-compatible endpoint through LiteLlm, whose parameter is api_base — not base_url.

ADK has no native OpenAI connector. It reaches one through its LiteLLM wrapper: LiteLlm, parameter api_base. Note the spelling — it is not base_url.

pip install google-adk "litellm>=1.84"

from google.adk.agents import LlmAgent
from google.adk.models.lite_llm import LiteLlm

model = LiteLlm(
    model="openai/zen5",
    api_base="https://api.hanzo.ai/v1",
    api_key="sk-...",
)

agent = LlmAgent(model=model, name="hanzo_agent", instruction="Answer precisely.")

Lands on POST /v1/chat/completions.

The openai/ prefix is LiteLLM's provider form and is what selects the OpenAI-compatible path; the segment after it is the model id. LiteLlm also takes extra_headers and extra_body. ADK requires litellm>=1.84, and LiteLLM's own docs carry a security advisory against versions 1.82.7 and 1.82.8 — pin above them.

MCP

McpToolset goes straight into the agent's tools list. For a remote server the URL is the url parameter of StreamableHTTPConnectionParams.

from google.adk.agents import LlmAgent
from google.adk.tools.mcp_tool import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StreamableHTTPConnectionParams

toolset = McpToolset(
    connection_params=StreamableHTTPConnectionParams(
        url="https://api.hanzo.ai/v1/mcp",
        headers={
            "Authorization": "Bearer sk-...",
            "Content-Type": "application/json",
            "Accept": "application/json, text/event-stream",
        },
    )
)

agent = LlmAgent(model=model, name="hanzo_agent", instruction="Answer precisely.", tools=[toolset])

StdioConnectionParams and SseConnectionParams are the siblings, and tool_filter narrows what the agent sees.

MCP

On this page