What is MCP? And why should we care?

Hello, fellow AI enthusiasts. It's been some time since I last wrote here, and for good reason: I've been swamped with work and life. But in the meantime, I've been investing my time in learning more about AI development and real-world applications, and it's been incredibly rewarding.

There has been a topic that has been buzzing my news feed constantly: MCP.

Imagine if every app on your phone had a different charging port. Frustrating, right? That's essentially what I discovered AI developers have been dealing with when connecting their systems to various services. Until now.

Since November, when Anthropic released something called the Model Context Protocol (MCP). Even OpenAI jumped on board. You might have heard about it too.

But here's where my confusion started.

My Journey to Understanding MCP

I'll be honest: I read Anthropic's official blog post twice. I asked both ChatGPT and Claude to explain MCP to me in simple terms. And even then, I couldn't grasp what MCP actually did in practice.

The reason? MCP itself does nothing that end users can see. If you're just using AI applications, you'll never even know MCP exists. It's completely invisible.

After a lot of back and forth, trial and error, and probably more coffee than I should admit, I think I finally cracked it. Let me save you the headache I went through.

What MCP Actually Does

Think of MCP as a universal translator between AI systems and the tools they need to access. Before I understood this, I was getting lost in all the technical jargon. Here's the simple version that finally clicked for me:

Before MCP, if an AI wanted to connect to Slack, Google Drive, GitHub, or your calendar, developers had to build custom integrations for each service. Every platform had its own unique way of sharing data. I spent hours trying to understand different API documentations, and trust me, it's not fun.

MCP changes this by creating a standardized layer between AI systems and external services. It's like having one universal remote that works with every device in your house, instead of juggling five different remotes and constantly losing them.

Here's what this means for different people (and yes, I had to figure this out for each group separately).

First, let’s visualize what changes before and after MCP:

If you're an AI enthusiast: MCP helps AI applications become more powerful when accessing your data across different platforms. Your AI assistant can seamlessly pull information from multiple sources without the developer needing to reinvent the wheel each time.

If you're a developer: This is where MCP really shines. It eliminates the headache I mentioned about building custom integrations. Once a service supports MCP, connecting to it becomes plug-and-play. More time for actual features, less time wrestling with documentation.

If you're a business owner: MCP opens doors to more sophisticated AI implementations. Real-time data access across departments, better insights, improved automation.

The Struggles I Discovered

Let me paint you a picture of what I learned about the pre-MCP world. Connecting an AI system to business tools looked like a spider web of custom integrations. Each connection needed its own maintenance, security setup, and troubleshooting process. When something broke, developers had to dig into service-specific code to fix it.

I actually tried building a simple integration myself to understand this better. What should have been a straightforward task turned into a weekend project that taught me why developers were so excited about MCP.

With MCP, this becomes much cleaner. Services that adopt MCP provide a standardized interface, making it easier to build robust applications. The benefits I discovered:

• Enhanced AI performance through better data access

• Reduced development time (and my weekend stress)

• Improved security through standardized protocols

• Easier scaling across different systems

• Greater flexibility in choosing tools

My Honest Take: Should You Jump In?

Here's where I might surprise you: not yet.

I know this sounds contradictory after explaining all the benefits, but let me share what I learned from digging deeper. Even Andrew Ng, one of the biggest names in AI, recently mentioned at a LangChain event that he's seen many MCP cases that are still confusing, with MCP servers that don't work correctly. While I think MCP is fantastic for serving data to AI applications, I also believe it might be too early to invest heavily in it.

This hit home for me because during my research, I encountered several MCP implementations that were buggy or poorly documented. The technology is definitely promising, but the execution isn't quite there yet.

If you're running a small or medium-sized business where every resource counts, the investment in switching to MCP might not pay off immediately. I learned this the hard way when I calculated the development time versus the current benefits.

However, don't ignore MCP entirely. Keep building with traditional methods, but watch for signs that the MCP ecosystem is maturing. That's exactly what I'm doing.

For Fellow Developers: My Learning Path

If you do have the resources to experiment (and more patience than I initially had), here's what I wish someone had told me upfront:

MCP follows a client-host-server architecture. Think traditional APIs, but with more structure. You have MCP clients, hosts (like Claude Desktop or IDEs), and servers that expose resources and tools.

The host manages multiple client connections, while each client maintains isolated connections to specific services. Servers expose the actual resources, tools, and prompts that AI can use.

You can implement MCP in two ways:

1. Build a service that provides MCP access to your data

2. Use existing MCP services in your applications

Both approaches use the MCP Python SDK, which honestly made my life much easier.

After trying multiple tutorials that left me more confused, I found Dave Ebbelaar's MCP crash course on YouTube. It's the most practical one I encountered, covering both server and client development with real examples. The GitHub repository is solid too. This is the tutorial that finally made everything click for me.

What I'm Watching For

MCP represents a significant step forward, even if most people will never see it working. It's like plumbing: you don't think about it until it makes your life better or worse.

Based on my research and experimentation, here's what I'm watching for before recommending full adoption:

• More major platforms offering native MCP support

• Better tooling and documentation

• Success stories from businesses similar to yours

• Stable, well-maintained MCP servers for common services

My Bottom Line

The protocol is heading in the right direction, but like many emerging technologies I've encountered, it needs time to mature. I'm keeping an eye on it while continuing to build with traditional methods.

My advice? Keep learning, keep building, and stay curious about MCP. The universal remote for AI systems is coming, and when it arrives, we'll want to be ready. But we don't need to rush into it just yet.

PS: If you have a team of three or more non-ai-focused developers and want to level up their AI knowledge to stay ahead of the competition, and increase their perceived job security, join our AI Agents Course waiting list.