What is MCP, or Model Context Protocol?
Most people describe MCP as tools that AI can use. Most people stop there.
Go one layer deeper and MCP is an interface for connecting AI to other platforms. Both explanations are useful. Neither is the whole thing. There are many ways to explain MCP because each one describes a different layer of the same thing.
Officially, MCP servers can expose tools, resources, and prompts to an AI application. Tools are simply the part most people notice first.
But what if MCP is actually intent designed to be repeated?
Break it down to first principles and you find some interesting ideas that most people may never get to. Start designing MCP servers and suddenly your eyes open.
Anything can become an interface
Most people latch onto the idea that if something has an API, it can be an MCP. Almost. More accurately, if software has an API, you can often build an MCP server in front of it.
Now AI can help map software into MCP tools by reading APIs, CLIs, documentation, schemas, and interface metadata. In practice, that means even proprietary software can sometimes be given an MCP interface, as long as you have a legitimate way to access or automate it.
I saw this in practice with a Samsung TV experiment. The television was proprietary hardware, but the controls and metadata it already exposed could still be mapped into tools an AI could understand. Samsung did not have to build an integration for my AI platform. The interface could be built around what was already there.
What does that mean for the future of software?
Far more software can be integrated into an AI platform than its creator explicitly designed for.
I don’t know. I’m just playing in this sandbox.
MCP is not the only integration path. You also have CLIs, skills, browser automation, and APIs. There is overlap, and the overlap is interesting, because AI now has multiple ways to use a platform even when that platform does not offer an official AI integration.
You can make an interface for almost anything. In some cases, your AI could create an account, configure it, and restore your preferred settings. Whether it should, and whether a service permits it, are separate questions.
The dead internet may look different than we thought
This brings me back to something I have thought about before.
The future may be a version of the dead internet, but not in the way people usually mean it. Each person may interact with the internet through their AI. That AI will have to parse all the AI-generated bot slop being thrown into the world.
Essentially, AI becomes a personalized interface for the internet.
This is not entirely hypothetical. According to Imperva’s 2026 Bad Bot Report, automated traffic accounted for more than 53 percent of all web traffic in 2025, up from 51 percent the year before. Human traffic fell to 47 percent.
That does not prove bot traffic is exponential, and not all bots are malicious. It does show that machines already represent the majority of traffic in that dataset.
This is why AI may become necessary for both users and companies. AI has gotten smarter at automating slop. Every day, someone finds a new way to automate more of it.
Here is the thing: this has always been happening. Bots are not new. What changed is that the tools are now available to the masses.
If automated traffic continues to grow, one of the main jobs of our models and internet interfaces will be parsing the slop bombarding the internet.
That is a problem.
MCP may be the solution and the problem
MCP may be part of the solution and, simultaneously, part of the problem.
A trusted, first-party MCP server can protect against slop by giving an AI direct, structured access to a source. Instead of scraping ten SEO pages, the AI can query the actual system.
On the other side of the coin, MCP can also make slop easier to reproduce.
Trash in, trash out, as they say.
You can download an MCP server, connect a collection of tools for posting to social media, and tell a bot to go. There are already systems designed to research successful YouTube channels, generate similar content at scale, and slowly work their way toward an audience and some ad revenue.
What this means for software development
If this continues, future software may need two things.
First, it needs a reason for people to come back. Good code is not enough by itself. Is there a hook? Is there a dopamine hit? Is there something people intentionally want?
Second, it needs a reason for AI to rely on it. Does it provide trusted information, useful actions, or access that an AI cannot easily replace?
That carves out the future of technology pretty simply, and it is a little scary.
What happens when future models can generate a AAA-level video game that took someone one week to put together?
Good code will still matter, but it may no longer be the moat. Distribution, trust, and being found will matter even more. If creators cannot earn those things, they may simply fade into history.
Your AI may become your UI
If any of this is true, advertising as we know it is also in trouble.
Carefully crafted funnels designed for human eyes become less valuable when users delegate browsing to their AI. Money spent on ads that an AI filters out is wasted money.
I do not think video will be immune.
As a proof of concept, I made a nested MCP server that lets ChatGPT open a YouTube video and play it inside ChatGPT itself. Soon, an agent may be able to watch the ads for you, skip or summarize the interruptions, and serve the actual video on a platter.
Whether a platform permits that behavior is a separate question.
That does not necessarily mean every advertisement disappears. It means the economics and placement of advertising may have to change. The valuable position may no longer be the page a person visits. It may be the source their AI trusts and recommends.
Soon, your chosen AI may be your user interface for the internet.
MCP, CLIs, APIs, skills, and browser automation do not prove that future, but they all point in the same direction. Major AI companies are building models that can understand your intent and act across software on your behalf.
MCP may help clean up the mess.
It may also make the mess easier to scale.
Probably both.