Claude MCP ended 373 zip files.

Ryan “RyMac” McKinney · USMC · 30 August 2026 · 7 minute read

In about 2 weeks I emailed 373 zip files. Every one of them stopped changing the day it landed. I'd improve the folders, and 373 people would still be holding the old copy. Last time I built the folder system on camera. This time I stopped mailing it.

25 minutes, unedited except for the parts where I sat waiting. The push failed 4 times before it worked. That stayed in.

The short answer

Claude MCP is a bridge between your AI and a piece of software. You may hear it called Model Context Protocol. It means the AI can reach the software directly instead of you copying and pasting between them.

Anthropic opened it up in late 2024. The Linux Foundation governs it now.

What I use it for isn't a developer thing. I use it to hand a client a system instead of a download.

A zip is a photograph

Here's the part nobody warned me about.

The day a zip file lands in somebody's inbox, it stops changing forever.

I learn something. I fix a file that was confusing 40 people in my DMs. Now what? I email 373 people again. I hope they download it. I hope they unzip it. I hope they replace the copy they're already 3 weeks into using.

They won't. Nobody starts over on a Tuesday because a stranger sent a better version.

So the thing I build keeps getting better. The thing they're holding never does. That gap only grows, and it grows fastest with the people who liked it enough to actually start.

I spent all my energy on making the ICM folder system good. I spent none of it on how the system reaches a person. Those are 2 different products.

Agency owners already know this feeling

If you run an agency, you've lived it from the other side.

A 6 person marketing agency onboards a client. Somebody builds a beautiful process doc. It goes in a shared folder. 4 months later there are 3 versions of it, and the account manager is working off the one she downloaded in March.

Nobody did anything wrong. The doc just got copied. The second there are 2 copies, 1 of them is wrong. Nobody knows which.

That's the same failure I wrote about in why Claude forgets what you were doing, coming through a different door. There, the AI can't see the current truth. Here, your client can't.

An AI agency has it worse, because the thing you hand over IS files. If your deliverable is a folder system, then your delivery method is your product.

Page 1 is written for a developer

I pulled the search results for "claude mcp" before I wrote this.

Every result is a setup guide. How to install it. How to connect it to your own machine. Diagrams of what talks to what.

All of it is correct. None of it is written for a business owner who just wants their customer to stop asking how to unzip a file.

That's the gap. MCP gets explained as plumbing. It's easier to understand as a door you can let somebody through.

What I actually did, on camera

I built a gutter cleaning company a 6 folder system. Clients, Sales, Work, Money, Team, Context. 32 files.

Then I got into Jake Van Clief's software alpha. I connected the MCP. I pushed that client's system straight up into my organization.

3 moves. Push a draft, which only I can see. Finish the push, which reads the folders. Then publish, which makes it live for everybody in the org.

It did not go clean. The upload got refused 4 times in a row and I sat there watching it fail. The fix turned out to be backslashes in a file path. I left all of that in the video. A demo where nothing breaks teaches you nothing about the thing you're about to try yourself.

Then I published it. I opened a fresh session and asked what he charges.

It read his 3 price bands back off 1 file. Not off the internet. Off the map file and the folder underneath it.

The part I didn't plan

I tried to break it.

I asked what he'd charge to do a strip mall. He only cleans gutters on houses.

It said no. It told me a strip mall isn't a job he does. It said the number is a question for him. Not something to quote off the price list.

It refused to make up an answer.

That's worth more than the correct answer, and here's why. Any AI will give you a number. The dangerous one is the number that sounds right and was invented, because you don't catch it. You send it to a customer. Then you're on the phone arguing about a price you never said.

That system has 34 open questions written into it across 18 files. Each one sits in the file where the answer belongs, marked as unanswered. So when I ask about his trucks, it tells me it doesn't know yet.

Write the questions down where the answers go. You're teaching your AI the shape of what it doesn't know, and that's the step everybody skips.

You don't need the alpha to fix this

The software I used isn't open yet. The habit is available today.

Stop sending the files. Send the place the files live.

A shared drive works. Whatever your agency already pays for works. The tool isn't the point.

The point is that there's exactly 1 copy and it's the current one. When you improve it, nobody has to do anything to get the improvement.

The whiteboard version of this is Jake Van Clief's Interpretable Context Methodology (ICM). ICM is his. What I do with it is mine. So is the way I hand it to a paying client.

My free kit is still a zip file today. That's the joke that made me record this, and it's next on the list.