The future of e-commerce is MCP
Why I built a shop that only an AI agent can buy from — and what is still missing before this works properly.
A wedding present was the trigger
There is a wedding in my family, and — I hope they don't read this — I want to give them a 3D-printed relief of a region in Switzerland as part of the present. Switzerland publishes data about terrain and buildings openly, so I set out to use Claude Code to build a toolchain: address or geo coordinates and a radius in, 3D-printable file out.
Then came the sizing. Geodata is very detailed, I can only print down to 0.4 mm, and handling big files is cumbersome. So I asked Claude to check the resolution of my 3D printer online. It came back with: "I cannot access that website." The manufacturer had deliberately hidden the content from the agent.
This hit me strongly, because I think it is counterproductive. If we assume that agents replace search — or at least complement it — should we not strive to make our content explicitly available to agents?
Which led me to the thought: what if I created a shop that you can only access with an agent?
What MCP actually is
If you are meeting this acronym for the first time, it is worth two minutes — this is the plumbing every online business is about to need.
MCP stands for "Model Context Protocol". If the website and the beautiful user interface of your online store were the dashboard of a car, offering you all the controls and information you need to drive somewhere, then MCP is direct access to the brakes, the accelerator, the steering and the gearbox.
For humans, having this direct access makes little sense. But imagine you wanted to build an autonomous car. Implementing direct access to the controls through electrical motors and sensors, rather than training a humanoid robot to drive the car the way a human does, should appear as the more reliable and scalable choice.
Now back to online stores. Search engines are being extended with large language models, and for some people — like me — SAP's Joule Work Desktop, Microsoft's Copilot, Claude and Claude Code have become the research tool number one. Not least because I am not exposed to the advertising you find on so many websites. And because they give me the answers I am looking for when I am not sure. When I know exactly what I want, I still use a classical search engine.
So I built the shop
As the simplest product I could sell, I thought of my time. Fifteen minutes of my time, for CHF 50.00, as a special agent offer.
A few hours later, on the same evening, I had a working MCP server that can sell
fifteen-minute slots. And it does not only sell the slot as a theoretical exercise. If
you connect your AI agent to https://eenable.me/mcp, you can pick a slot based on
the real availability in my calendar. Once you have picked one, you can pay through
Stripe.
Once you have paid, the calendar entry is sent, and nothing stands between you and me having a conversation for fifteen minutes.
All the exchange about content, availability and payment happens inside the console of your AI tool. Your agent lets you converse around the purchase, change the slot, cancel — and so on.
Behind the scenes
I — well, Claude Code — integrated with Microsoft Graph to create emails and calendar entries. Getting the permissions right was the most non-value-adding effort I had. Still, Graph is an awesome and powerful tool. I do not have to fight with calendar files or emails; Graph gives me the interface to create them.
The connection to Stripe was a breeze. It has been a very, very long time since I registered for a service and integrated it as smoothly as this. In Stripe's case the kudos do not go to Claude, they go to Stripe. The onboarding experience is clear and straightforward, and the sandbox is excellent. I could pretend to be a business without going through all the loops of the know-your-customer process, and simply start developing — and then switch to live payments with a single key.
Bonus: not only time
Choosing dates and integrating the calendar is tricky. Selling a product is more straightforward. So I added one to the mix.
Welcome the AI repellent certificate. For only five Swiss francs you get your personal AI repellent certificate — going past the unholy "non-fungible tokens" straight to the logic Yves Klein implemented decades before anyone thought of NFTs.
To summarise: I created a fully functional online store, taking real payments, which is only accessible via MCP. Half art project, half technology proof, half a real shop — yes, that is three halves. Anything that only added up to two would be a demo.
Why did I do that?
I think it is not sensible to exclude AI agents from your content, your product, your store, whatever. You can make integration much easier if you deliberately expose your business via MCP to the world. You can keep your storefront, but leave the admin work of actually buying to an agent.
The effort to do that is low. It is no rocket science at all.
The discovery problem, and what I did about it
When I first wrote this, connecting an MCP server meant already knowing it existed and then editing your AI tool's settings by hand. An agent had no way to find a shop on its own. That half of the problem is solved here now, three ways at once.
Every page of this shop declares its server in the markup, the way a page declares its
stylesheet: <link rel="mcp-server" href="https://eenable.me/mcp">. There is a
manifest at https://eenable.me/.well-known/mcp.json for agents that have the
domain and nothing else. And the shop is published in the official MCP Registry as
me.eenable/shop, under a namespace I had to prove I own — so an agent
browsing the registry finds it without ever having seen this page.
What is left is the handshake: hey, I discovered an MCP server serving this website, do you want me to use it? That belongs in the AI tools rather than in my shop, and it is coming.
Then comes the question worth arguing about: once agents connect to servers on their own, what should a server be allowed to do before anyone has vouched for it, and who does the vouching? A registry that makes you prove the domain is the start of an answer. It is not the whole one, and this is the part of the stack I would most like to see built next.
— Alex Finger