Coffee Shop

A hands-on Model Context Protocol (MCP) project. An LLM (VS Code Copilot in Agent mode) takes your coffee order and "makes" it by coordinating a Barista server and four machine servers.

Built with the official MCP Python SDK's FastMCP.

---

How it works

The LLM is the orchestrator. Servers are dumb specialists — none of them talk to each other. The Barista returns a recipe, and the LLM walks that recipe across the machines.

flowchart TD
    User([You]) --> LLM[VS Code Copilot<br/>orchestrator]
    LLM --> Barista[Barista server<br/>menu · orders · recipes]
    LLM --> Grinder[Grinder]
    LLM --> Brew[Brew unit]
    LLM --> Steamer[Steamer]
    LLM --> Dispenser[Dispenser]

---

Order flow

sequenceDiagram
    participant U as You
    participant L as Copilot (LLM)
    participant B as Barista
    participant M as Machines

    U->>L: What's on the menu?
    L->>B: get_menu()
    B-->>L: 4 drinks
    U->>L: Large latte, extra shot
    L->>B: place_order(...)
    B-->>L: order id + recipe
    L->>M: grind → brew → steam → dispense
    L->>B: mark_order_ready()
    L-->>U: Your latte is ready ☕

---

Menu

Drink	Milk?	Notes
Espresso	No	Base shot
Americano	No	Espresso + hot water
Latte	Yes	Steamed milk, light foam
Cappuccino	Yes	Steamed milk, thick foam

Machines

Component	Job	Used by
Grinder	Beans → grounds	All
Brew unit	Pull the shot (+ Americano water)	All
Steamer	Texture milk	Latte, Cappuccino
Dispenser	Assemble the cup	All

Espresso skips the Steamer. Latte vs Cappuccino differ only in foam thickness.

---

Project layout

coffee-shop-mcp/
├── .vscode/mcp.json
└── src/coffee_shop_mcp/
    ├── server.py       # Barista
    ├── grinder.py
    ├── brew_unit.py
    ├── steamer.py
    └── dispenser.py

---

Setup

uv venv
uv add "mcp[cli]"

Test one server in the browser Inspector:

uv run mcp dev src/coffee_shop_mcp/server.py

---

Run in VS Code

.vscode/mcp.json:

{
  "servers": {
    "coffee-shop": { "type": "stdio", "command": "uv",
      "args": ["run", "python", "src/coffee_shop_mcp/server.py"] },
    "grinder":     { "type": "stdio", "command": "uv",
      "args": ["run", "python", "src/coffee_shop_mcp/grinder.py"] },
    "brew-unit":   { "type": "stdio", "command": "uv",
      "args": ["run", "python", "src/coffee_shop_mcp/brew_unit.py"] },
    "steamer":     { "type": "stdio", "command": "uv",
      "args": ["run", "python", "src/coffee_shop_mcp/steamer.py"] },
    "dispenser":   { "type": "stdio", "command": "uv",
      "args": ["run", "python", "src/coffee_shop_mcp/dispenser.py"] }
  }
}

1. Open the folder in VS Code, click Start on each server in mcp.json.
2. Open Copilot Chat → Agent mode.
3. Say: "What's on the menu? Then make me a large latte and run it on the machines."

---

Notes

• In-memory only — orders reset when the server restarts.
• Simulated hardware — machines return text results, nothing physical happens.
• Idle servers get stopped/restarted by VS Code automatically — that's normal.

A learning project. ☕