MCP client

Client is the part of an application that speaks MCP to one server, and it can connect to a server object in memory as this course does.

Last updated: 29 Sep, 2026 · MCP 2.2

The last lesson listed a tool. Now call it. The same Client that lists tools also runs them, over the real protocol, whether the server is in memory, a subprocess or across a network.

The Client API

from mcp import Client

async with Client(mcp) as client:      # a server object means in memory
    await client.list_tools()          # what the server offers
    await client.call_tool(name, args) # run one tool

Connecting in memory

Client(mcp) is given the server object, so it connects in memory: no process to start, no port. async with connects when the block starts and disconnects when it ends. Every client method is async, so it is awaited inside main, which asyncio.run starts.

async with Client(mcp) as client:
    result = await client.call_tool("lookup_order", {"order_id": "A17"})

from mcp.server import MCPServer

mcp = MCPServer("Shop support")

ORDERS = {
    "A17": {"item": "blue mug", "status": "shipped", "total": 12.50},
    "B42": {"item": "desk lamp", "status": "processing", "total": 48.00},
}


@mcp.tool()
def lookup_order(order_id: str) -> str:
    """Look up an order by its id and say where it is."""
    order = ORDERS[order_id]
    return f"Order {order_id}: {order['item']}, {order['status']}."

Calling a tool end to end

The whole call in one file. The result comes back in three parts.

import asyncio

from mcp import Client
from shop import mcp


async def main():
    async with Client(mcp) as client:
        result = await client.call_tool("lookup_order", {"order_id": "A17"})
        print(result.content[0].text)
        print(result.structured_content)
        print(result.is_error)


asyncio.run(main())

Order A17: blue mug, shipped.
{'result': 'Order A17: blue mug, shipped.'}
False

The three parts of a result

• content is a list of blocks for the model to read; here, one text block.
• structured_content is the same result as data for the application's code; a string return is wrapped as {"result": ...}.
• is_error says whether the call failed, and Tool errors covers when it is True.

Reading what the connection knows

When a client connects, the two sides agree on a protocol version, and the server declares its capabilities: the kinds of request it will answer.

import asyncio

from mcp import Client
from shop import mcp


async def main():
    async with Client(mcp) as client:
        print(client.server_info.name)
        print(client.protocol_version)
        print(client.server_capabilities.tools)
        print(client.server_capabilities.completions)


asyncio.run(main())

Shop support
2026-07-28
list_changed=True
None

MCPServer always declares tools; Resources and Prompts arrive later in the course. Completions, argument autocomplete, needs a handler this server does not have, so it is None, and a client will not ask for it.

In memory vs a subprocess vs HTTP

When to use the in-memory client

• Trying a server while you build it, with no process to launch.
• Testing a server, where the in-memory client becomes a pytest fixture in part 6.

Related

• Previous: MCPServer quickstart
• Next: MCP Inspector with mcp dev
• Reference: MCP Python SDK client