MCP 서버란 무엇인가 (그리고 실제로 무슨 일이 일어나는가)

Source: Dev.to

Overview

So far we’ve covered:

• why MCP exists
• what MCP is
• what tools are

Now let’s answer a key question: When the model decides to use a tool… who actually runs it?

What an MCP Server Is

• The component that exposes tools and executes them.
• Not just your backend – it sits on top of your backend and is designed specifically for LLM interaction.

Responsibilities of the MCP Server

1. Describe available tools to the model

   • What tools exist
   • What they do
   • What inputs they need

   This is the information the model “sees”.

2. Validate requests before execution

   • Are required fields present?
   • Are types correct?

   Validation prevents bad or unsafe execution.

3. Perform the real work

   • Database queries
   • API calls
   • Business logic

Simple Flow Example

User request

Show my last 3 orders

Model output (structured)

{
  "tool": "get_user_orders",
  "arguments": {
    "user_id": "123",
    "limit": 3
  }
}

MCP server processing

• Checks that user_id is present and limit is a number.
• Executes something like:

SELECT * FROM orders WHERE user_id = 123 LIMIT 3;

• Sends the structured data back to the model.

Model final response

Here are your last 3 orders…

What the Model Never Does

• Touches your database
• Calls APIs directly
• Runs any code

The MCP server handles all of that.

Benefits of This Separation

• Better security – the model never accesses resources directly.
• Cleaner architecture – clear division between decision‑making and execution.
• Controlled execution – validation and sandboxing happen in the server.
• Reusable systems – tools can be shared across models.

Analogy: Restaurant

• Model → decides what to order.
• MCP server → kitchen that prepares it.
• Tools → items on the menu.

The model doesn’t cook; it only selects.

LLM‑Friendly Tool Design

LLMs need:

• Clear, descriptive names
• Structured inputs (e.g., JSON)
• Simple actions

Poorly designed tools can lead to crashes, incorrect actions, or security issues.

What the Model Does Not Do

• Decide which tool to use (the model selects, but the server executes).
• Understand user intent beyond tool selection.
• Generate the final response (the model does this after receiving tool results).

Foundation of MCP

• Model decides which tool to use.
• Server executes the tool.

Now that we understand the execution layer, who connects the model and the server?

That’s the MCP client – the component that makes everything work together.