BOTTOM LINE
Most people have never heard of the Model Context Protocol (MCP), but it’s quickly becoming one of the most important pieces of AI infrastructure, with monthly downloads up almost 4,750 percent in under two years. Instead of forcing AI agents to click through websites like a human, MCP gives them a secure, standardized way to interact directly with software. That’s why companies like Google, Microsoft, Amazon, Salesforce, and thousands of others are racing to build connectors. They’d rather hand an AI agent the keys than let it climb through the window.
The AI Story Most People Aren’t Talking About
When people talk about AI agents, the conversation usually centers on what they’ll be able to do.
Book flights.
Write code.
Answer customer support tickets.
Schedule meetings.
Those are the visible applications. Underneath them is a much quieter story that’s getting far less attention: how does an AI agent actually use software? That question has become one of the biggest infrastructure races in AI.
The technology driving it is something called the Model Context Protocol, or MCP. Most people have never heard of it.
Within about sixteen months, nearly every major AI company had adopted it.
From Experiment to Industry Standard
Anthropic introduced MCP in late 2024 as a standard way for AI models to interact with external tools and data. It didn’t remain an Anthropic project for very long.
OpenAI added support. Microsoft integrated it into Copilot Studio. Google adopted it through Vertex AI. Amazon followed with Bedrock.
By this year, monthly SDK downloads had exploded from roughly two million to nearly one hundred million: a jump of almost 4,750 percent in sixteen months. More than 5,800 MCP servers now connect AI models to everything from developer tools and business software to e-commerce platforms and productivity applications.
That’s remarkable growth for something most people outside the AI industry have never heard of.
Two Ways an AI Agent Can Use Software
There are really two ways to let an AI agent interact with software.
The first is what most people have seen in demonstrations. The AI watches a screen the way a human does. It clicks buttons, types into boxes, scrolls through web pages. In theory, that means it can use almost any application without the software company doing anything.
The downside? It’s guessing. Even the best computer-use models only succeed on about three out of four standardized desktop and web tasks, and change the interface to something unfamiliar, that number drops further.
The second approach is completely different. Instead of watching the screen, the AI connects through a defined interface the software company intentionally built. That’s what MCP provides.
Google’s Gmail connector is a great example. Rather than clicking through an inbox, the AI simply asks Gmail to search messages, retrieve conversations, create drafts, or apply labels. Only the functions Google explicitly exposes are available. Nothing more, nothing less.
I think the difference comes down to this: a computer-use agent is a guest fumbling around looking for the light switch.
An MCP connector is the homeowner handing over a labeled set of keys. Those keys only open specific doors.
A browser session doesn’t work that way. Log an agent into a full browser, and it technically has the run of the house: delete the inbox, change a password, wander into an unrelated tab. MCP never hands over that kind of access in the first place.
That’s not a guarantee against everything. But it’s a fundamentally smaller blast radius, by design.
Why Companies Are Building Connectors
Building an MCP connector takes engineering time, and companies wouldn’t invest in it unless there was a clear payoff. There is: the biggest advantage is control. Instead of letting an AI click around their software however it wants, companies define exactly what an agent can read, what actions it can take, and what stays off limits. That creates a safer, more reliable experience for both customers and the software provider.
There’s another reason. Businesses are beginning to expect AI integration: analysts expect roughly 65 percent of enterprises to require it, in some form, by the end of this year. If you’re buying enterprise software today, one of the questions is increasingly becoming: can it work with AI agents?
But I think the most interesting reason is also the simplest. AI agents are going to use popular software whether companies prepare for them or not. The only real question is how they get inside.
A company that builds an MCP connector decides exactly what an AI agent can and can’t do. A company that doesn’t build one leaves that decision to a browser automation model clicking through its interface instead. A much less controlled outcome.
The Infrastructure Race
Most headlines about AI focus on bigger models and smarter chatbots. I think the more interesting race is happening underneath them: the companies building the infrastructure often end up becoming the platforms everyone else depends on.
We’ve seen that before. The internet gave us browsers. Cloud computing gave us APIs. Mobile computing gave us app stores.
AI agents are creating something new: standardized doors into software. That might not sound as exciting as a breakthrough model announcement. Infrastructure rarely does.
But history shows that the companies controlling the infrastructure often end up shaping everything built on top of it.
