MCP: Connecting AI and Websites
So the website should be able to talk to AI now? Then there is barely a way around MCP. But a word of caution: what hides behind those three letters can mean quite different things.
Barely an AI conversation gets by without MCP. Sooner or later the acronym inevitably lands on the desk of the people running a website: do we actually need this too? We at F7 have been building websites since 1996, and we have seen a fair few supposedly revolutionary standards come and go again. MCP might stick around. But right now a lot of nonsense is being spread about it, too.
More interesting than retelling the protocol from A to Z are three questions:
- Which variants are there?
- Which of them is relevant for the website in question?
- And what should be decided before the technology sets the direction?
What is MCP, exactly?
MCP stands for Model Context Protocol. The name sounds technical; the idea behind it is fairly simple. It is an open standard for how AI assistants talk to external systems — which data and actions those systems offer, and in what form. MCP was introduced in late 2024 by Anthropic; by now it is supported by practically every major AI assistant.
The usual explanatory image is the USB-C connector. We prefer one from everyday life: the menu in a restaurant. Every kitchen cooks differently, but a menu always works on the same principle. The guest does not need to know the kitchen in order to order something. MCP is the menu of a given system: it describes what the AI can request — in a form that different assistants understand. What goes on the menu, however, is decided by the kitchen, not by the guest.
Communication and permission
And now the important part: MCP is not an AI model, and not a feature that can simply be added on. What it does is make sure systems understand one another.
Who is then allowed to do what is a different question altogether. And these two things are exactly what tends to get lumped together in the current MCP enthusiasm.
Welche Arten von MCP gibt es?
Which kinds of MCP are there?
There are three variants. Two of them are interesting for a website — and in quite different ways:
- AI tools accessing external services
- The system itself becoming an MCP server
- WebMCP: the public website becoming operable for AI agents
1. AI tools accessing external services
An AI assistant connects via MCP to SEO tools, analytics or a calendar. It can then query data or trigger actions without constant jumping back and forth between five tools.
Useful? Absolutely. A matter for the website? No. The website is a bystander here.
2. The system itself becoming an MCP server
This is where it gets interesting for websites: the content management system provides a menu of its own. Creating pages, editing text, uploading images — an AI can handle all of it on behalf of the editorial team.
For TYPO3, such a server already exists as an open source project. This is no longer just "AI can read content". This is AI working inside the system.
And this is precisely where a technical interface suddenly becomes an organisational question:
- What is the AI allowed to do?
- What is it not allowed to do?
- And who approves its changes?
That calls for a clean permissions and approval concept.
3. WebMCP: the public website becoming operable for AI agents
WebMCP is a proposal from Google and Microsoft, developed in a W3C community group — a draft, not yet a standard. It concerns the browser side: the website tells agents which functions it offers, such as booking a table or filling in a form. This is outward-facing, and still one for the future — but it is already being confused with the second variant.
Kurzübersicht
| Variant | Who connects to whom | Affects | Decision |
| 1. External services | AI tool → third-party provider | Workflows | Which tools do we use? |
| 2. Own MCP server | AI assistant → own CMS | Editorial team | Who may do what, and where is it approved? |
| 3. WebMCP | AI agent → public website | Customers | What do we expose to the outside? |
MCP in editorial practice
What changes for the editorial team once MCP is connected?
The editorial team writes the brief, the AI implements it in the CMS, a human approves. In concrete terms: editors describe to the assistant what should go on the page. The AI creates the content, but it does not push it live: in TYPO3 it lands in a workspace, a draft area. Only after review is it published. Best of all, workspaces have been part of the TYPO3 core for years — no third-party extension required.
We run exactly this setup for an international hospitality client, with real editors and real day-to-day business.
To be honest: not everything automatically gets faster. Approval takes time, and somebody has to read what the AI has written. The gain lies elsewhere: the AI takes over the legwork, the human stays in control.
The permissions concept is what determines security
Because MCP does not give AI any magical special privileges. It gets exactly the permissions it is granted. If a system is properly secured, there is no need to invent a new security framework for MCP – established permission and access-control concepts can be used instead.
The main risks arise when AI is connected to other systems: credentials or digital keys can be stolen, interfaces can be manipulated, or the AI can be prompted into taking unintended actions. The more permissions it has, the greater the risk.
Recommendation
The sensible way in is the second variant, small and controlled: a dedicated account for the AI, a workspace, a clear approval step. WebMCP belongs on the watch list, but not yet on the roadmap.
Where we can help:
- Assessment: with the F7 Expertcheck we review the roles, content structure and approval paths of a TYPO3 installation.
- Integration: API-Integrations – from the MCP server in the CMS through to specialist systems.
- Editorial: Content management – roles, workflows and training for the team.
Fazit
MCP is not a feature to be bought in — it is a decision about how processes can be improved. Because something shifts here: the human no longer operates the CMS directly. They tell the AI what should happen, and the AI works inside the system. Anyone who has settled that shift organisationally can implement MCP within weeks and benefit straight away.
We are happy to look at your setup and tell you where you stand: get in touch!
Weiterführende Links
- Offizielle Spezifikation – Primary source for the standard
- WebMCP – Draft Community Group Report – W3C Web Machine Learning Community Group, draft specification for the browser
- Standard statt Sonderlösungen – Mittelstand-Digital Zentrum Chemnitz
- MCP: Die Sicherheitsrisiken des Model Context Protocols (german)
Author
