Practical MCP with FastMCP & LangChain

Understanding How the MCP Client Works

Let's see, step by step, how the MCP client works and how it connects the OpenAI model to the MCP server.

Step 1: Establishing the Connection

The code starts by opening a "pipe" to the server using the streamable HTTP transport.

async with streamable_http_client(url) as (read, write, _):
    async with ClientSession(read, write) as session:
        await session.initialize()

The initialize() step is the initial handshake. It works like dialing a phone number where our client and the server "say hi" to each other and agree on which version of the protocol they are speaking.

Step 2: Discovery and Translation

The LLM (GPT-5-mini) has no idea what tools our server has. The client has to "introduce" them using 2 steps:

• Step 1 - Discovery: session.list_tools() asks the server, "What can you do?"

• Step 2 - Translation: The code loops through the results and converts them into the specific format OpenAI understands (Function Calling format). It maps MCP's inputSchema directly to OpenAI's parameters.

Step 3: The First Question

The client sends the question of the user and the list of translated tools to OpenAI and asks: "This is the question and these are the tools you can use, how would you answer the question?"

response = openai.chat.completions.create(
    model=model, messages=messages, tools=tools
)

At this point, OpenAI doesn't give an answer yet. Instead, it looks at the question, looks at the tools, and says: "I need to use tool 'X' with arguments 'Y' to answer this."