jd five

What Is MCP

So far, tools lived inside your agent file.

Teams often keep tools in another program: orders DB helper, notes app, company API. You need a shared protocol:

  1. What tools exist?
  2. Call this tool with these inputs
  3. Here is the result

MCP (Model Context Protocol) is that open protocol.

Real shape (stdio)

Piece Role
Server Separate process that owns tools
Client Starts the server, speaks MCP, lists/calls tools
Transport Often stdio — stdin/stdout as the wire

Do not print() debug lines on the server’s stdout. That wire is for MCP messages.

Client process talks to a server process.
  mcp_client_demo.py          orders_server.py
  (client)                    (FastMCP server)
       │  spawn + stdio              │
       ├────── list_tools ──────────►│
       │◄───── tool list ────────────┤
       ├────── call get_order ──────►│
       │◄───── order facts ──────────┤

ShopTiny use case

Support needs order facts without importing shop code into every agent.

Tool Job
get_order Look up id → item, amount, status
list_orders List known order ids

Data lives next to the server (data/orders.json). The client only speaks MCP.

This chapter vs the next two

Chapter You learn
What Is MCP (here) Client lists + calls a real server
Build an MCP Server You author tools + policy data
Agents with MCP Tools An LLM agent uses MCP tools in a loop

Language note

Official MCP SDKs are strongest in Python and TypeScript — those are the Code labs.
Java and Go tabs stay empty here on purpose (no hand-rolled stand-in). Switch to Python or TypeScript for the real framework path.

See it in Code

Install once: pip install mcp (Python) or the TS SDK note in Code.
Run the client demo — it starts orders_server.py, lists tools, calls get_order("4412").