What Is an MCP and Why Should I Care

MCP changed how AI tools connect to real systems. If you build with AI, this is the protocol to understand.

I know I say “this changes everything” a lot lately. I’m saying it more and more because, honestly, everything I’m doing is changing everything.

I was a 2000s IT guy who moved into management. Now I’m a 2020s guy reevaluating my entire toolset. I feel like I’ve been building barns with wooden nails while the rest of the world moved to pneumatic nail guns. I managed the guys using the nail guns. I just haven’t spent a lot of time learning what a nail gun is.

That’s changing. Fast.

In a recent post I mentioned Claude Code, an AI that runs in a terminal on a dedicated machine inside my home lab stack. Depending on what permissions I give it, Claude can access my network. Would I do this at work? Maybe. After a long pause. Only inside a self-contained VLAN with test data and disposable Proxmox VMs where I can see what it does before it touches anything real.

At home? I’m a little more aggressive. If I get hit by a bus, my home network stops. But it’s my network, nobody else depends on it. At work, if I’m the only one who builds something and I disappear? Yeah, people give a damn. Different risk tolerance. Different rules.

I wanted an MCP server, a Model Context Protocol server. Big fancy words for a structured way for an AI to make API calls to other systems. I knew I wanted Claude to be able to do things for me, not just answer questions.

So I did what I always do. I researched. For hours. Tried to understand what MCP is. Tried to figure out how to build one myself. Started a project plan. Made it exceptionally complicated.

Existing MCP servers you can download are just someone else’s version of which API calls are available for a specific tool. Building one from scratch means getting all the API calls formatted correctly in JSON. It’s doable, but it’s work.

Then I asked Claude a simple question: “Can you help me build an MCP server?”

In perfect Claude fashion, we’d been working together for a few weeks at this point,the response was something like: “Yes. Or would you like me to just build one?”

Excuse me. You can just build one?

Those are actual words I typed into the prompt. Yes, Claude said. If you want an MCP server to interact with your UniFi controller, I can build one.

I said: but the problem is I might want to write to the UniFi controller, and the official MCP server from Ubiquiti only has read-only commands.

“That is correct. Would you like me to build one with full or limited read-write commands?”

The light bulb turned into a lightning bolt.

I don’t have to research for hours to build an MCP server. I can tell Claude what I want it to do, that’s the guardrail. Claude can then build the server and execute those tasks for me.

That’s it. That’s the whole insight. I define the boundaries. The AI builds the tools. And then the AI uses the tools.

I’m not going to pretend I fully understand where this goes yet. But I know this: my workflow just changed permanently. Stay tuned. This is going to get interesting.

Leave a Reply

Your email address will not be published. Required fields are marked *