What Is MCP? Why Can’t an AI Just Read a File on My Own Machine? | MCP Part 1

2026年8月14日

Run an AI on your own machine for a while and the thought arrives: can this thing not just read a file sitting on the same computer? It answers questions, but it cannot read a file. It cannot open a browser. It cannot make an image. It never gets outside the conversation.

MCP (Model Context Protocol) is the mechanism for getting past that wall. The name tells you nothing on its own — “a protocol for handing context to a model" — but what is underneath is simple, and using it made it click for me.

So what is MCP, in the end? And why can an AI not read a file as things stand? I will go through it while showing how I actually use it.

This is where things stood in July 2026. For the specification I am referring to the latest revision, 2026-07-28.

What was the AI unable to do?

An LLM (large language model — the part of an AI that handles text) is, put crudely, a component that takes text and returns more text. Anything outside the text it was handed is, in principle, out of reach.

If you wanted it to read a file on your machine, a human had to copy the contents and paste them in. Ask about the weather and you got whatever was true when it was trained. Ask for an image and there was nothing there to draw with; the AI has no such function of its own.

For a while, every company solved this “cannot reach outside" problem in its own way. The mechanism for calling tools differed from product to product, and an extension built for one AI did not run on another. From the builder’s side, that presumably meant rebuilding the same thing once per AI you wanted to support.

What did MCP actually decide?

What MCP settled is how an AI and an external function talk to each other. Anthropic published it in 2024, and since then it has been run as a shared standard with several companies involved rather than as one company’s property.

What happens when the way of talking is shared? An extension you build once runs on any AI that supports it. The same story as the shape of a USB connector being fixed, so any manufacturer’s mouse fits any computer. You often see MCP described as “USB for AI," and I think the analogy holds up well.

The word “server" makes people brace themselves, but in most cases it is a small program running on your own computer. It is not something you rent from anywhere. Part 3 goes into this properly.

What is actually happening, seen on my own screen

Explanation alone is hard to hold on to, so here is what I use day to day for work on this blog. All of it runs on a desktop PC (Ryzen 9 3950X / RTX 3090) and on a GMKtec EVO-X2 (from here on, EVO-X2) mini PC.

What using it taught me is that adding MCP does not make the AI smarter. What grows is its reach. Same model, wider range of things it can do.

An MCP server carries three kinds of tools

In the specification, what an MCP server can offer an AI splits into three. Knowing the names makes the descriptions of published servers much easier to read.

The one to watch is tools. Unlike resources, which only read, these change something. Install a server carrying a tool that deletes files and the AI can delete files. The usefulness and the danger come from exactly the same place.

Can you use MCP with a local LLM?

This is probably the part readers of this blog care about. The answer was “yes, but it depends on the model."

To use MCP, the AI has to decide “a tool should be called here" and then call it in the prescribed format. That ability depends on the size of the model and what it was trained on, and the smaller the model, the shakier it gets. Even in what I tried here, I saw models answer in plain prose when they should have reached for a tool, and get the call format wrong.

Frontier cloud models are comfortable with tools; models at a size that runs on your desk still seem to be some way behind. That is not the same as “unusable" — cutting down the number of tools and being specific in your instructions made it workable in some situations. Part 3 comes back to this.

Knowing the spec has just changed a lot will save you confusion

While I was writing this, the MCP specification was revised: the 2026-07-28 version. It is described as the largest change since release, and reading it, the foundations really have been swapped out. So that nobody new to this gets thrown by an older explanation, here are the points.

What changed is that staying connected is no longer assumed. Until now you exchanged an opening greeting and then carried on down the same line, like a phone call. In the new specification the greeting is gone, and each request goes out like a letter carrying whatever information it needs. The term for this is stateless.

That said, the exchange has not disappeared wholesale. A new query, server/discover, is now required, letting you ask in one go what the other side supports. It is closer to say that the procedure did not vanish; it moved.

And one important thing: the three players and the three kinds of tools above are unchanged by this revision. There is nothing extra for a newcomer to learn, so no need to worry on that front.

There is one change that does catch people running AI locally, though: sampling has been deprecated. That was the mechanism letting an MCP server ask “may I borrow your AI for a moment," and the official migration path given is “connect directly to your LLM provider’s API." The route for letting a server use the model on your own machine has narrowed. It is not disappearing immediately — deprecated features are specified to remain for at least 12 months.

Wrap-up: MCP widens reach, not intelligence

What to take from Part 1:

• MCP is a standard that fixed a shared way of talking between an AI and external functions. Build an extension once and it runs on any AI that supports it
• The only thing you choose is the MCP server. Most are small programs running on your own computer
• What MCP raises is reach, not how clever the model is
• The 2026-07-28 version dropped the always-connected model (it went stateless). The three players and the three kinds of tools did not change

So how many MCP servers are there really, and which ones should you install? Next time I count the official registry and sort them by what you would use them for.

Sources

Checked 30 July 2026. Specification referenced: 2026-07-28 (current).