DSPy

dspy.LM takes api_base. Note the spelling — it is not base_url. The parameter is not in the explicit signature; it lands in **kwargs, which DSPy forwards to LiteLLM, and the API reference names it as an LM field.

pip install dspy

import dspy

lm = dspy.LM(
    "openai/zen5",
    api_base="https://api.hanzo.ai/v1",
    api_key="sk-...",
    model_type="chat",
)
dspy.configure(lm=lm)

qa = dspy.Predict("question -> answer")
print(qa(question="What moves the tides?").answer)

Lands on POST /v1/chat/completions.

The openai/ prefix is LiteLLM's provider form; the segment after it is the model id. model_type takes chat, text or responses, so POST /v1/responses is reachable from the same constructor.

MCP

DSPy converts MCP tools rather than owning a client. dspy.Tool.from_mcp_tool(client, tool) turns each one into a dspy.Tool you hand to a module such as dspy.ReAct, so the server URL is configured on the MCP SDK client, not on any DSPy parameter.

import dspy

tools = [dspy.Tool.from_mcp_tool(session, tool) for tool in (await session.list_tools()).tools]
react = dspy.ReAct("question -> answer", tools=tools)

session is a ClientSession from the MCP Python SDK, connected however that SDK connects — for https://api.hanzo.ai/v1/mcp that is its streamable HTTP client, with the bearer header set there.