agenticsearch
Search DE, Deutsch

Model Context Protocol

A plug for language models.

MCP governs how an AI system addresses tools and data sources. For companies it is less a question of visibility than one of connection.

Romana Hieß holding a small robot in her palm, Florian Hieß holding its remote control

Definition

What the protocol governs.

The Model Context Protocol is an open standard through which AI applications reach tools, data sources and templates without building a separate connection for every pairing. Anthropic published it in 2024, and several providers support it by now.

Before it, every AI application needed its own connection to every system. With five applications and five systems that is 25 connections. With a shared standard it is ten.

In practice MCP consists of a server offering capabilities and a client inside the AI system using them. The server states what it can do, the model decides whether it needs that.

Relevance

When this is a topic for you.

For visibility in AI answers, MCP is not the deciding factor. Anyone wanting to be named in ChatGPT needs a retrievable website, not an interface. That is the most common confusion in this field.

It becomes relevant in two cases. First internally: when staff are to use an AI tool that reaches your ERP system, your CRM or your documentation, MCP is the tidy route there.

Second externally: when agents are to act with you, so not only read but query availability or book. That is still the exception today and will become the rule in some sectors.

Components

What an MCP server offers.

Capabilities standard
  • Tools functions the model may call, such as an availability query
  • Resources data the model may read, such as documents or records
  • Prompts prepared sequences an application can offer

Granting permissions stays your job. The protocol carries requests, it does not decide who may do what.

Position

What we make of it.

The standard solves a real problem and is therefore not a passing fashion. At the same time we currently see many offerings selling MCP as a visibility solution, and it is not that.

Our order of work: first make the website retrievable and unambiguous, then measure, and only afterwards think about interfaces. Anyone starting at step three is building a door into a wall nobody is standing at.

When a connection is due, it belongs to the AI brain work, because the same question sits behind it: which information is machine readable, current and unambiguous?

Common questions

About the Model Context Protocol.

What does MCP stand for?
Model Context Protocol. It describes how an AI application addresses tools and data, independently of which model sits behind it.
Do we need an MCP server to be named in ChatGPT?
No. Mentions come from retrievable content, not from interfaces. Anyone selling it otherwise is conflating two things.
Is MCP tied to one provider?
The standard comes from Anthropic and is open. Several providers support it, which is what makes it useful in the first place.
What does such a connection cost?
That depends entirely on the system being connected. We quote no range for it, because a figure without knowing your systems would be guesswork.
How secure is it?
As secure as you build it. The protocol governs transport, not authorisation. Permissions, logging and limits stay your job.
Can a model then reach everything?
Only what the server offers. That is why choosing the offered capabilities is the actual security decision.
Is this meant for internal tools or for customers?
Both are possible, internal is far more common today. External access by third-party agents is still the exception.
Does MCP replace our API?
No, it usually sits on top of one. Anyone with a clean interface has done most of the work already.
Do you offer MCP development?
We advise on it within the AI brain work and implement the connection with your developers. We do not sell it as a product of its own.
How do you spot offerings that promise too much?
By the claim that an MCP server improves visibility in AI answers. There is no mechanism and no evidence for that.

Next step

Let us find out where you stand in AI answers.

One conversation, 30 minutes, no sales pressure. You leave with an honest read on whether this is worth the effort for you.

Look at the audit first Request a call

Free first analysis

We run one question through several AI systems and send you the result. No subscription, no sales call required.

The address this is about.

The question you want to show up in. For example: “Which agency helps with AI visibility in Austria?”

This is where the result goes.

Usually answered within two working days.

Search

    All pages at a glance