Multi-Turn Tool Loop

Closing the loop from Lesson 15

Lesson 15 got as far as the model deciding to call add. Now we actually run it and hand the result back, the same three-step loop from LangChain's tool-calling lesson, just with an MCP-backed tool instead of a local one.

messages = [HumanMessage("What is 12 plus 30?")]
response = model.invoke(messages)
messages.append(response)

for call in response.tool_calls:
    tool = tools_by_name[call["name"]]
    result = await tool.ainvoke(call["args"])
    messages.append(ToolMessage(content=str(result), tool_call_id=call["id"]))

final = model.invoke(messages)

ToolMessage carries the result back with tool_call_id matching the original call, so the model can line up which result answers which request, identical to how the LangChain course closed its loop.

A detail specific to MultiServerMCPClient

Watch the logs while this runs: you'll see the server's tools/list handler fire more than once. MultiServerMCPClient opens a fresh connection to the server for each tool call, rather than holding one session open across the whole conversation. For a stdio server, that means a new subprocess launch per call. This is a real cost (slower than a persistent connection) that the adapter accepts in exchange for never needing you to manage a session's lifetime by hand, fine for a calculator, worth knowing about before you build something call-heavy.

Multiple turns, one conversation

The messages list is the same running conversation history from LangChain's conversation memory lesson: each turn appends to it rather than starting fresh, so a second question can refer back to the first one's answer.

async def ask(model, tools_by_name, messages, question):
    messages.append(HumanMessage(question))
    response = model.invoke(messages)
    messages.append(response)

    for call in response.tool_calls:
        tool = tools_by_name[call["name"]]
        result = await tool.ainvoke(call["args"])
        messages.append(ToolMessage(content=str(result), tool_call_id=call["id"]))

    if response.tool_calls:
        response = model.invoke(messages)
        messages.append(response)

    return response.content