Connecting to Multiple Servers

Why this matters

Lesson 1 introduced the host/client picture with several clients, one per server, feeding into a single host. Every lesson since has used exactly one server. Real hosts, Claude Desktop, Claude Code, rarely stop at one, a typical setup connects a filesystem server, a search server, and a company-specific server all at once. This lesson is where MultiServerMCPClient earns its name: it already accepted a dict of servers back in Lesson 15, we just only ever put one entry in it.

Two servers, one client

client = MultiServerMCPClient(
    {
        "calculator": {
            "transport": "stdio",
            "command": sys.executable,
            "args": [str(calculator_server_script)],
        },
        "weather": {
            "transport": "stdio",
            "command": sys.executable,
            "args": [str(weather_server_script)],
        },
    }
)
tools = await client.get_tools()

get_tools() now returns tools from both servers, flattened into one list. The model doesn't need to know or care which server a given tool came from, bind_tools(tools) works exactly as before, one flat list, same as Lesson 15.

Naming collisions

Two servers could, in principle, both expose a tool called search. get_tools() accepts a server_name argument to fetch tools from just one server if you need to disambiguate, and MultiServerMCPClient's tool_name_prefix option can prefix every tool name with its server's label automatically. This lesson's two servers use distinct, specific names (Lesson 4's advice), so no collision happens here, but it's worth knowing the escape hatch exists once you're combining servers you don't control.

model = ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite").bind_tools(tools)
messages = [HumanMessage("What's 12 plus 30, and what's the weather in Seattle?")]
response = model.invoke(messages)

for call in response.tool_calls:
    print(call["name"], call["args"])
    # add({'a': 12, 'b': 30})
    # get_weather({'city': 'Seattle'})