The Website for Agents: MCP Servers

Earlier this week, I shared my favorite comparison for understanding the state and role of MCP in the landscape of building AI tools and systems.

In case you missed that go back and read “It's time to critique MCP”.

But the key comparison is this:

Browsers are to Websites, as Agents are to MCP Servers.

I find this comparison to be really valuable in a lot of ways, including when we think about the optimizations that we’re seeing agents make in their use of MCP Servers like I talked about in the previous article.

The Universal Question: Where Should I Focus?

As a career web developer, I get to focus more on building the best websites and apps I can build by leaving things like browser optimizations up to the browser implementers.

In the same way, I’m leaving AI agent optimizations up to agent implementers, and focusing my efforts on building the best MCP servers I can.

That’s why all of the workshops I’ve created so far are about building with MCP servers, not clients!

It’s a perfect parallel, really. Similar to how MANY more people build websites than build web browsers, my bet is that many more people will build MCP servers than clients.

Much like web standards played a huge role in allowing countless people to build their websites without worrying about which browser their users choose, MCP is a new standard that offers the similar possibilities of building an integration once that works anywhere the MCP standard is supported.

And the number of users using MCP-implementing software is rapidly increasing.

Standards are the key to building useful abstractions

When Cloudflare published their “Code Mode” blog post, I asked myself the question many others asked:

“If LLMs are so good at writing code, what use is MCP? Can’t people just publish their REST API with documentation and have the LLM write directly to those?”

Basically, if a coding agent can write a bespoke bit of code to integrate with an API, couldn’t that deliver on the promise of MCP without the MCP?

But the truth that anyone who’s been building software with an AI assistant knows is that LLMs need a lot more hand-holding for that approach to be reliable, let alone flexible enough to work across multiple agent clients and platforms.

And it makes sense! Humans struggle to navigate a comprehensive API, even with a bunch of great documentation on the endpoints. There are just so many variables and scenarios to account for.

Great documentation tells you how it works, but that’s not the same as it telling you how to use it to solve your specific problem or scenario. Add in important factors like backward compatibility and complexity like data migrations over time.

See how quickly an API-only approach can become confusing, for developers and AI assistants alike?

Turns out things that help humans often also help AI models.

The good news is that we already have a solution for this kind of problem: abstraction.

I think about it this way: what is a website if not an abstraction on top of internal APIs that are designed to handle specific use cases for our users?

That’s exactly how I see an MCP server, as an abstraction on top of internal APIs that is designed for an AI agent to press the buttons instead of a human user.

Instead of handing the AI agent a bunch of individual APIs, you start with an intentional set of design decisions. You decide on a core set of tools and resources that an LLM can stitch together on-demand in lots of ways, with a high degree of certainty that things will “just work.”

And because the structure of those tools and resources are standardized, anyone building AI agents can connect to any MCP server that provides tools and resources.

The best part? Your LLM can interact with the MCP server directly with tool calling… the application can implement something like “Code Mode” and have the LLM write code for interacting with the tools. None of these optimizations change how you build your MCP server.

The key lesson: these optimizations become possible and universally applicable when they’re built on top of a standard like MCP.

So I’m going to keep building MCP servers and teaching you to do the same.

AI agent builders are going to keep making their agents use those servers more efficiently.

And as long as we stay focused on the right parts, the user experience across the board will continue to get better and better, and allow agents to do more and more for them.

And I’m so stoked about that future and helping get you there!

— Kent

P.S. Backed by popular demand, I’ll be launching the Master MCP Cohort Workshops as a self-paced edition with end of year pricing in early December. So now’s a good time to start looking at your end of year training budget and be ready to sign up during the early bird discount period.