The Ultimate Guide to Model Context Protocol, Part 2: Behind The Scenes of MCP

In our previous post, we introduced the Model Context Protocol (MCP) and how it transforms our dear Claude from a knowledgeable yet impotent AI into a helpful digital butler who can actually interact with your files, apps, and services. Now it’s time to draw back the curtain and have a gander at the magic behind it.

Don’t worry—we’ll keep things simple and jargon-free, dishing out plenty of analogies to explain the technical concepts like my Uncle dishes out expletives when India cocks up a cricket match. By the end of this post, you’ll understand what makes MCP tick and how you can start exploring different MCP servers for your specific needs.

How MCP Works

Remember our analogy of MCP as a universal translator between AI and your digital world? Let’s expand on that to understand what’s happening behind the scenes.

The MCP Architecture Explained

At its core, MCP follows what tech folks call a “client-server architecture.” This is exactly how computers work with the internet. The browser on your computer is the “client”. It retrieves and displays information from a “server” over the internet via a protocol called HTTP.

The Model Context Protocol is similar. Let’s say you’re enjoying a cold Sunday evening by the fire in the study of your manor, as one does. It’s a high-tech study with a built-in AI assistant. You ask the assistant to have some hot toddy sent over:

1. The Host (where you chat with AI)

• This is an application like Claude Desktop or Cursor where you interact with an AI assistant.
• In our analogy, this is the study of your manor.

2. The MCP Client (the translator)

• This is built into the host application. So the engineers at Claude and Cursor need to build this first for the whole thing to work.
• It translates between what the AI understands and what MCP servers speak.
• You never interact with this directly, it works behind the scenes like HTTP.
• In our analogy, it’s an upgrade module for your study that allows your AI assistant to communicate with other parts of your manor, such as the bar.

3. MCP Servers (specialized helpers)

• Each server is like a specialist with access to specific resources.
• One server might know how to work with files, another with Slack, and so on.
• Servers can be on your computer or connect to online services.
• In our analogy, the bartender who makes the hot toddy and brings it over to you is the server.

4. Tools (actions your AI takes via servers)

• These are the functions available to the AI on the server.
• A document server may have a read_file action that the AI can invoke to read a specific file.
• In our analogy, the tool is the ability to prepare libation.

5. Resources (your digital stuff)

• The actual files, apps, and services the AI needs to access
• Could be local (on your computer) or remote (on the internet)
• In our analogy, these are the ingredients that go into making the hot toddy. I prefer a spot of Cognac myself.

If you enjoyed this analogy, I have more for you. Be a dear and sign up to my newsletter for more.

A Day in the Life of an MCP Request

Ok enough with the analogies. To really understand how this works, let’s follow what happens when you ask your AI assistant to summarize a document and send it to Slack:

1. You make a request to Claude: “Be a good sport and summarize the quarterly_report.pdf on my desktop. Oh and while you’re at it, post the key points to the #team-updates Slack channel”
2. Claude recognizes this requires access to both files and Slack, so it needs to use MCP
3. The MCP Client activates and connects to two different MCP servers:
   • The File System MCP Server (to access the PDF)
   • The Slack MCP Server (to post the message)
4. Permissions are checked:
   • The File System server asks: “Allow Claude to read quarterly_report.pdf?”
   • The Slack server asks: “Allow Claude to post to #team-updates?”
   • You approve both requests
5. The File System server retrieves the PDF content and sends it back through MCP
6. Claude processes the document and creates a summary
7. The Slack server takes Claude’s summary and posts it to your team channel
8. You receive confirmation that the task is complete

All of this happens in seconds, with the complex technical work hidden from view. The beauty of MCP is that it handles all the complicated connections while maintaining security and giving you control.

The Technology That Powers MCP

Now that you understand the basic flow, let’s demystify some of the technology that makes MCP possible:

The Protocol Itself

The Model Context Protocol is what tech people call an “open standard.” This means:

• It’s publicly documented so anyone can build with it
• It follows consistent rules for communication
• It’s designed to be secure from the ground up

Think of it like the rules of the road—all vehicles (or in this case, different software) follow the same rules, allowing smooth traffic flow.

Security Measures

MCP takes security seriously with several built-in protections:

Permission-Based Access

• Nothing happens without your explicit approval
• Permissions are fine-grained (specific to each action)

Sandboxing

• Each MCP server is isolated from others
• If one server has a problem, it doesn’t affect the rest

Audit Trails

• All actions are logged so you can see what happened
• Useful for troubleshooting or monitoring usage

Real-Time Communication

MCP uses modern, efficient methods for passing information back and forth:

• It’s designed for low latency (minimal delays)
• It handles both simple requests and large data transfers
• It manages two-way communication seamlessly

This means you don’t have to wait long for results, even when dealing with complex tasks involving multiple systems.

MCP Servers: The Building Blocks of AI Integration

MCP servers are the workhorses of the system. Each one is specialized for a specific purpose, and you can mix and match them based on your needs.

Types of MCP Servers

MCP servers generally fall into a few categories:

1. Local Resource Servers

• Access things on your computer
• Examples: File System, Local Database, Browser Control

2. Communication Servers

• Connect to messaging and social platforms
• Examples: Slack, Email, Bluesky

3. Productivity Servers

• Integrate with work tools
• Examples: GitHub, Google Drive, Calendar

4. Information Servers

• Fetch and process data
• Examples: Weather, Search, Wikipedia

5. Specialized Servers

• Handle niche needs
• Examples: 3D Printer Control, Smart Home

Where to Find MCP Servers

In the previous post, I mentioned a few of the top MCP servers. If you’re looking for more, there are several places to discover and download MCP servers:

1. Official MCP Servers Repository
2. MCP.so Directory
3. Glama.ai
4. Composeio
5. Cursor Directory
6. Awesome MCP Servers
7. Fleur MCP App Store
8. MCP Run
9. Smithery

Setting Up Your Own MCP Server

While most people will simply use existing MCP servers, you might be curious about how they’re created. Or perhaps you can’t find one and want to build your own. Here’s a simplified explanation:

What You Need to Create an MCP Server

If you’re not a developer, you probably won’t be creating your own MCP servers. But understanding what goes into them can help you appreciate what they do:

1. Programming Skills

• Knowledge of languages like Python and JavaScript
• Understanding of APIs and web services

2. Development Tools

• MCP SDK (Software Development Kit)
• Required libraries and dependencies

3. Access to Resources

• API keys for external services
• Documentation for the systems you’re connecting to

For the Technically Curious: A Simple Example

Here’s what a very basic MCP server might look like in concept (this is simplified pseudocode):

// Define what the server can do
server.addCapability("read-weather-forecast", {
  description: "Gets the weather forecast for a location",
  parameters: {
    location: "The city or area to get the forecast for",
    days: "Number of days to forecast"
  },
  securityLevel: "requires-approval"
});

// Implement the actual functionality
server.onRequest("read-weather-forecast", async (request) => {
  // Get the forecast from a weather service
  const forecast = await weatherAPI.getForecast(
    request.parameters.location, 
    request.parameters.days
  );
  
  // Return the results
  return {
    current: forecast.current,
    daily: forecast.daily,
    warnings: forecast.alerts
  };
});

// Start listening for connections
server.start();

This simplified example shows how an MCP server:

1. Defines what capabilities it offers
2. Specifies what parameters are needed
3. Sets security requirements
4. Implements the actual functionality
5. Returns results in a structured format

In reality, MCP servers are more complex, with proper error handling, security features, and optimization—but this gives you a sense of their basic structure.

Connecting Multiple MCP Servers: The Power of Combination

One of the most powerful aspects of MCP is the ability to use multiple servers together. This creates workflows that would otherwise require complex programming.

Example: A Research Assistant Workflow

Imagine you’re researching a topic and want AI help. With multiple MCP servers, you could:

1. Use the File System server to scan your existing notes
2. Use the Browser Control server to search for new information
3. Use the Wikipedia server to verify facts and get background
4. Use the Google Drive server to save your findings
5. Use the Slack server to share insights with colleagues

All of this could be accomplished with a single request to your AI assistant, with each server handling its specialized part of the task.

Common Questions About MCP Servers

“Are MCP servers safe to install?”

MCP servers from reputable sources follow strict security protocols. Stick to official directories and well-reviewed options. Each server will ask for specific permissions, so you always maintain control over what they can access.

“How many servers should I install?”

Start with just the ones you need for your common tasks. You can always add more later. Most users begin with the File System server and add others as needed.

“Will MCP servers slow down my computer?”

Most MCP servers use minimal resources when idle and are designed to be efficient. If you’re not actively using them with your AI assistant, they have very little impact on performance. I’ve noticed, however, that it does slow down my Claude Desktop app if I add too many.

“Can I use MCP servers with any AI assistant?”

Currently, MCP works with compatible hosts like Claude Desktop and Cursor. As the protocol gains popularity, more AI applications are likely to support it.

What’s Next on Your MCP Journey

Now that you understand how MCP works behind the scenes and what servers are available, you’re ready to start building your personalized AI workspace.

In my next post in the series, I’ll provide a hands-on guide to building out useful agentic workflows with Claude and MCP servers. I’ll walk through the setup process with screenshots and troubleshooting tips to ensure a smooth experience.

Sign up below and stay tuned for it!