A 1980 personal computer and a modern laptop facing each other across a brass threshold line on a workshop bench

mcp praxiswissen mittelstand agentic-ai

Your data is the asset. The agent is the excuse.

The infrastructure layer of MCP has already won. The mass-market moment has not arrived. That gap is exactly where the cheap experiments live.

6-min read

Table of Contents
· 7 min · 1,242 words

There is a moment every technology has, somewhere between the lab and the living room.

The personal computer had it around 1979. The machine had existed for years as a hobbyist toy. Then VisiCalc shipped for the Apple II at $99.95, sold roughly 700,000 copies over its life, and people started buying the computer just to run the spreadsheet. The technology historian Steven Levy put it plainly: after that, the computer was a business tool, and everything changed.

MCP is somewhere in that gap right now. The question worth asking is which side of it you want to be on.

The plumbing has already won

Let me be concrete, because the numbers matter here.

Anthropic introduced the Model Context Protocol in November 2024. By March 2026 the SDKs were being downloaded 97 million times a month, up from about a hundred thousand sixteen months earlier. The official registry holds more than 9,400 servers. Every lab that matters ships it: Anthropic, OpenAI, Google, Microsoft, AWS, Salesforce. In December 2025 Anthropic handed the protocol to the Linux Foundation, which is what you do when you want a standard to outlive you.

And in May 2026, SAP put 60 million euro into n8n at a 5.2 billion valuation, and started wiring it into Joule Studio. That is the signal for anyone in the DACH Mittelstand. Not a Californian lab announcement. A boring European enterprise vendor deciding the workflow layer is where the money goes.

So the protocol war is over. MCP won the way HTTP won. Quietly, underneath everything.

The mass-market moment has not

Here is the part the breathless coverage skips.

None of this means your bookkeeper is about to assemble her own agents. The five big MCP directories between them list tens of thousands of servers, and not one of them is usable by a normal person. They are npm, not the App Store. Gartner, for its part, reckons only 17% of organisations have actually deployed agents, and predicts more than 40% of agentic projects will be cancelled by the end of 2027. They have a word for the rest: agent washing.

A neat stack of timber, folded blueprints and hand tools laid out on a workshop surface, no finished structure in sight

This is normal. Mosaic shipped in 1993. Mass web adoption took until the very end of that decade. What closed the gap was never the protocol. It was the curation on top: the ISPs, the bundled browser, the Yahoo home page, the App Store editorial team. The thing that made it usable for everyone arrived years after the thing that made it possible.

So your instinct is right if you feel like you are being sold a house by being handed a pile of manuals and a stack of timber. Nobody became a carpenter by owning the tools.

But that is precisely the argument for moving now, not against it.

Why the gap is the opportunity

The reason to jump on the train while it is still pulling out of the station is not that agents will run your company next quarter. It is that the experiments are cheap right now, and they will not stay cheap.

Think about what you are actually buying when you connect your company to an LLM today. The chance to find out whether your data is clean. Whether it is transparent. Whether it is machine-readable at all. Most company data, in my experience, is none of those things. It lives in inboxes and someone’s head and a spreadsheet last touched in 2021.

Getting that data into reach of an agent forces you to confront all of it. That work pays off no matter what happens to any individual tool. Clean, structured, legible data is the asset. The agent is just the first thing that made you go and tidy it.

And the risk is more controllable than it looks. The depth of sharing is yours to set. Which systems get exposed and which stay behind the wall is yours to decide. Start with read-only access to one boring corner of the business and learn everything you need to.

Right now that experimentation is close to free. Your people can poke at it, find a venture you had not seen, analyse an opportunity nobody had time for. That permission slip costs almost nothing today. It will cost real budget the moment this stops being experimental and starts being procurement.

You do not have to build anything

The good news for a company under twenty people: the ecosystem is mature enough that you mostly do not have to build.

A single cable being connected between two plain wooden blocks, one hand making the one link, everything else empty space

The services you already pay for probably already speak the protocol. Zapier exposes 9,000-plus apps and 40,000-plus actions through a single MCP endpoint. If you can set up a Zap, you can set this up. Microsoft made it generally available inside Copilot Studio. The vertical software your industry already runs on is shipping agent features that are MCP underneath, whether the brochure says so or not.

Connecting any of it to Claude or to your editor is genuinely copy and paste. One URL. That part is easy.

The hard part is not the wiring. It is the data. Gathering it, cleaning it, deciding what is safe to expose. That is the boring work, and it is the work that actually matters, and no vendor can do it for you.

There is no single VisiCalc this time

People keep waiting for the one app that flips everything. I do not think it arrives in that shape.

VisiCalc worked because the spreadsheet was one idea every business shared. Agents are not one idea. The thing MCP actually changes is subtler. It moves us off the chat box, the DOS-prompt of this era, and toward something different for each person: a tool, a small visual, a workflow that fits one employee’s actual job. Your bookkeeper gets one surface. Your project lead gets another. Neither of them is typing into a blinking cursor.

That is the shift. Not one killer app, but the end of one-size interfaces.

At some point we will all have first-class models running inside the firewall, and the question of what you exposed and to whom will look quaint. But we are not there yet. Today the models live outside, the experiments are cheap, and the only thing standing between your company and the learning is the decision to start.

The protocol is laid. The train is at the platform. Getting your company on board is the one part nobody can do for you.

Sources

MAKE YOUR CASE.