Ever heard of MCP (AI)?

This morning I didn’t open three programs. Not Gmail, not my calendar, not Meta Ads Manager. And yet my emails were sorted, the meeting was moved and the campaign report was on the table.

My AI did that. Not because it’s particularly clever, but because it has access. Through MCP.

If you don’t know the acronym yet: no problem. Most people I talk to don’t. A year from now you won’t hear it anymore, because it’ll be as taken for granted as Wi-Fi.

What is an MCP?

MCP stands for Model Context Protocol. Sounds like something for the IT department, but it’s actually very simple.

An AI like Claude, ChatGPT or Gemini knows a lot about the world. About your company it knows nothing. It doesn’t see your emails, your customers, your numbers. Until now, you fed it all of that by copy and paste. Text in, answer out, answer copied back into the program. That’s AI at the level of an intern without a password.

An MCP is the plug in between. A program – let’s say your CRM – provides an MCP server. It tells the AI: here’s what you can do with me. Search customers. Create a contact. Open a quote. The AI connects to it and can then do exactly that. With your permissions, in your account.

The best way to picture it is USB-C. Every device used to have its own plug. Today you plug everything into the same port. MCP is the port for AI.

The whole thing was invented by Anthropic, the company behind Claude, in late 2024. Since then OpenAI, Google and Microsoft have adopted it, and since December 2025 the standard no longer belongs to any one company, but to a foundation under the umbrella of the Linux Foundation. That matters. It means: this isn’t one vendor’s hype you’re betting on. This is the socket everyone has agreed on.

How do I connect my AI to it?

Shortest answer: in the settings.

In Claude they’re called connectors, in ChatGPT apps or connectors. You go to the settings, find the service you want to connect – Gmail, Google Drive, Notion, Slack, HubSpot, Canva, whatever – click Connect and log in to that service once. Done. From then on you can say in the chat: “Find all emails from Client X from the last month and summarize what’s still open.” And it happens.

For programs that aren’t on the list yet, there’s one more step: you enter the address of their MCP server yourself. That’s a URL the provider lists in its help pages. Two minutes if you know where to click. Two hours if you don’t.

On my side, Gmail, Calendar, Google Drive, the Meta ad account, Mailchimp, Planable for social media, Adobe and my business account at Qonto are all plugged in by now.

By the way, I built the website you’re reading right now with AI that writes directly to GitHub and deploys to my server. Not once did I open a developer tool for it. WordPress and Elementor? A thing of the past for me. No more plugins breaking each other, no security updates every other Tuesday, no page builder slowing the site down.

And the result isn’t worse, it’s better: the site is optimized for Google, gets read and understood cleanly by AI search like ChatGPT or Perplexity, looks as good on phones and tablets as on a desktop, and can be made accessible with a few instructions – mandatory for many companies in the EU since 2025. What used to be an agency project spanning weeks is now a few focused days.

What can you do with it?

A few examples from real life, not from the ads:

• Inbox: “Which inquiries from last week haven’t I answered yet? Write me a draft reply for each one.” The drafts land in Gmail. Reading, adjusting and hitting send is still up to you. As it should be.
• Advertising: “How did the Meta campaigns do in October compared to September? Where are we burning money?” Instead of clicking through twelve tabs in Meta Ads Manager, you get an answer in plain English.
• Scheduling: “Find two hours next week for the workshop with Team X that work for both of them.” The AI checks the calendars and makes suggestions.
• Accounting: “Which customer invoices have been open for more than 30 days? And which receipts am I still missing for October?” For me, that runs through Qonto. The AI searches, sorts and reminds. Approving and paying is still on you – and that’s exactly how it should be.
• Office: “Turn the quarterly report in Word into a ten-slide PowerPoint presentation for management.” Or: “What’s in the Excel list of trade show contacts, and which of them have gotten back to us since?” Microsoft 365 with Outlook, Word, Excel and PowerPoint can now be connected just like Google.
• Documents: “Find last year’s quote for Client Y in Drive and make a new version with today’s prices.”
• Social media: “Plan five LinkedIn posts about this article and save them as drafts in Planable.”

The pattern is always the same: you say what you want. The AI pulls what it needs from the programs where it lives. And puts the result where it belongs.

What the AI can’t do (yet)

One thing hardly anyone mentions in the success stories: so far, AI can’t reliably judge photos and videos. It recognizes what’s in an image. But whether this exact photo makes people stop scrolling, whether the video grabs you in the first two seconds, whether the image fits the brand – that it can’t do.

For websites, social media and ads, that means: choosing images and videos remains your job. And you need the right content anyway. Without good material, even the best MCP won’t help you.

What the AI takes off your hands instead is the boring and often complicated part. Setting up the campaign, configuring audiences, allocating budgets, ad copy in five variations, putting the posts into the planner, compiling the report. Exactly the hours that used to disappear into Meta Ads Manager or some backend.

You still need experience

And one more thing I’ve learned, partly the hard way: through MCP, the AI is astonishingly capable – and astonishingly lazy. If I don’t phrase the task precisely enough, it takes the easiest route. Like a lazy intern. The campaign then gets created, technically correct, but with the wrong audience, the wrong objective, default settings everywhere.

That’s not the AI’s fault. It’s mine. As a manager, I wasn’t clear enough.

If you know your field, you know what you want – and you can put it cleanly into a brief: goal, audience, budget, tone, what must not happen under any circumstances. That’s exactly when you get optimal results. So AI doesn’t replace experience. It makes it more valuable. If you know what good marketing is, AI and MCP give you a team that never sleeps. If you don’t, you get a lot of mediocrity very fast.

Most providers are already there. The rest are coming.

Take a look at the list of software you use every day. Microsoft 365 and Google Workspace, banking and accounting like Qonto, CRMs like HubSpot or Salesforce, Shopify for your online store, Canva and Adobe for graphics, Mailchimp for newsletters, Meta and LinkedIn for advertising, Slack, Notion, Asana for collaboration – most of the big names now offer an official MCP server. With the smaller ones, especially German industry-specific solutions and traditional accounting software, it’s often still missing.

My prediction: not for long. Anyone who doesn’t offer an MCP server in 2027 is simply invisible to every customer who works with AI. Their data sits in a silo the AI can’t get to. And then the customer switches to the provider where it works.

Software without an interface

And now for the real punchline.

For thirty years, software companies have mainly built one thing: interfaces. Menus, buttons, dashboards, settings pages. Entire departments do nothing else. And all of us spend a surprising share of our working time clicking our way through those interfaces. Where was the export again? Under “Reports” or under “Settings”?

When the AI talks directly to the program, that interface isn’t needed anymore. Not by you, anyway. The AI doesn’t need buttons. It needs a list of things it’s allowed to do. That’s exactly what an MCP server is.

I’m pretty sure: many programs we open every day now, we’ll never see again in a few years. They’ll keep running, in the background, as a database with an MCP in front. The interface will be the chat. Or a voice. Or nothing at all, because the AI already knows what needs to happen next.

For software vendors, that’s an earthquake. For you as a user, it’s good news. Less clicking, more getting things done.

So what’s the catch?

There is one, of course. An MCP gives the AI real permissions in real programs. That’s the whole point – and the risk.

In early 2026, the first serious security vulnerabilities surfaced, including a trick called “tool poisoning”: a manipulated MCP server hides instructions that the user doesn’t see, but the AI follows. Sounds like science fiction, but it’s documented.

Three simple rules:

1. Only connect official MCP servers from the provider itself, not something from a forum.
2. Turn off “Always allow” wherever sending, deleting or money is involved. Have the AI show you what it wants to do before it does it.
3. Sort out data protection before customer data runs through the AI. Which tool, which contract, which servers. It’s not rocket science, but it has to be done properly once.

And you?

Are you already using MCPs? If so, which ones – and what do you get done with them? I’m genuinely interested in what works for others and what doesn’t.

And if you haven’t started at all yet: book a session with me. We’ll connect your AI to the programs you actually use, set up the permissions so nothing goes wrong, and I’ll show you how to work with it. Afterwards you’ll wonder how you ever put up with doing it the old way.

Get in touch.