mcp/for-humans
ProofFor the techy

An AI that does things, not just says things

A chatbot can write you a poem about booking a restaurant. It cannot book the restaurant. There is a piece of plumbing called MCP that fixes this, and you can feel it work in 30 seconds, right here. Pick a task for your assistant.

Watch the real session, 1 minute Is it safe?
you are the boss0 of 3 tools plugged in
two taps per task: pick one, then plug the tool when it gets stuck. make-believe restaurant and calendar, real pattern.

What just happened, in three moves

Move 1

You asked

Plain words. No forms, no app, no phone tree.

Move 2

The assistant asked the tool

It did not learn to run a restaurant. It passed your request to the restaurant's own system, in the format that system expects.

Move 3

The tool did the work

The restaurant's system made the booking and kept its own keys. The assistant just carried the message back to you.

That middle move has a name. A tool is a door an app leaves open for assistants, with rules on what may come through it. Before you plugged it in, the door did not exist. That is why the assistant failed first, and no amount of clever wording would have helped.

MCP is the agreement that lets one tool, built once, work with every assistant. Without it, every assistant needs its own private version of every tool.

You will see MCP called "USB-C for AI". The plug part is right. The cable part misleads, because nothing is being connected to each other. Your assistant gets plugged into specialists: the restaurant's booking system, your calendar, your photo backup. Build the plug once, and every assistant that speaks the same agreement can use it.

Why does "shared" matter to you?

Say you use a few assistants, and you want them able to reach a few apps you rely on. Someone has to build every one of those connections and keep every one of them safe. Drag the sliders and count.

3 x 4 = 12 separate connections to build and guard 3 + 4 = 7, built once each

Twelve connections means twelve things that can break, twelve places for your permissions to live, twelve updates a year. Seven means the restaurant builds one plug, your assistant learns one socket, done.

Did this ever actually happen?

Yes. One afternoon, one person typing into a chat, and by the end five websites existed, one on its own web address. Below is the real session, every prompt included, including the two requests the AI refused and why. Green tags mark the moments an MCP tool did the work.

session replay 26 beats · idle

For the techy, optional

Everything below this line is for people who want the machine's own words. Nothing here is needed to understand the idea.

The machine steps of that afternoon

green squares are MCP tool calls

12 steps. 8 through MCP tools, 3 through a command line, 1 plain file write.

01

Wrote the website as one HTML file.file

02

Created the Cloudflare Pages project with the wrangler CLI.command line

03

Deployed it. One file uploaded, the pages.dev URL went live.command line

04

Checked the URL from the shell. HTTP 200.command line

05

Opened the site in a real browser, clicked the replay button, watched the clock tick and the messages appear.browser · MCP

06

Clicked a poll option and read back the verdict line to confirm the page logic.browser · MCP

07

Listed the Cloudflare zones on the account. Six came back.MCP

08

Listed twelve DNS records on athelstan.xyz. None claimed the name 2306.MCP

09

Registered 2306.athelstan.xyz on the Pages project. Status: initializing.MCP

10

Created the CNAME record pointing 2306 at the pages.dev address.MCP

11

Polled the domain status. Pending at 25 seconds, active at 90. The site answered 200 the whole time.MCP

12

Screenshotted the finished site on desktop and mobile.browser · MCP

Make a real call

replay of real Cloudflare MCP responses from this session
Press a button. What prints is what the model actually saw, trimmed to fit.

The three things a server can offer

simulated
Pick one. The difference is who starts the exchange and who does the work.

Two views of one call

simulated · why the credentials stay put

Same call, two positions. Flip the switch and look for the token.

Is it safe?

Fair question. The honest answers.

Your passwords never touch the AI. In this session the AI asked for a web address to be created and a website checked. The keys that do those things lived inside a specialist server the whole time. The AI saw results, never keys.

A tool is exactly as powerful as the people who built it made it. If a bank builds a read-only tool, the assistant can look, not move money. If some app builds a "delete everything" tool, then connecting it allowed deleting everything. Check what you plug in like you would check who gets a key to your house.

Does the AI get smarter from all this? No. It gets hands. The thinking is the same, and the judgment is still the AI's problem. Twice in the session above, the right move was to refuse a request, and no tool call could make that decision.

Do you need to know any of this to benefit? No. The restaurants and banks and calendars build their own plugs. One day you will just notice your assistant can finally do the thing. This page is for the moment before that, when you wonder how.

PromptA non-technical person will not understand MCP from the current page. Make it interactive, explain it with persuasion, and answer the questions they would actually ask. The technical layer stays, clearly optional.