DEFINITION
What is MCP?
Model Context Protocol, the socket between your AI and your systems.
An agent that is not allowed to touch anything is a chat window with an expensive subscription. MCP is the standard that lets AI reach the software you already run, without commissioning something bespoke for every combination.
MCP, in full the Model Context Protocol, is an open standard that lets AI models and agents connect to external systems in a uniform way: databases, documents, APIs and tools. Instead of building a bespoke connector for every combination of model and system, everything speaks one language. Anthropic published the standard in November 2024, after which it was adopted by several providers. W69 AI Consultancy in Amsterdam connects AI agents through MCP to the systems organisations already run.
The arithmetic is not research but multiplication. Five agents each needing eight systems: without a standard that is 5 × 8 = 40 connectors to build and maintain. With MCP it is 5 + 8 = 13. At ten agents and twenty systems it becomes 200 against 30.
Why bespoke connectors catch up with you
The first connector is always fun. The fortieth is a department. Here is what happens when every link between an agent and a system is custom work.
Maintenance grows faster than the system
Every connector is code somebody wrote and somebody has to keep maintaining. When a system changes, the connector breaks. With five agents and eight systems that is forty places where something can quietly fail.
You get locked into your first choice
If your connectors are built around one model, switching is no longer a decision but a project. And models replace each other faster than your depreciation schedule. Build bespoke work on a vendor and you buy their pace along with it.
Nobody knows who can reach what
Forty connectors means forty separate occasions on which somebody configured permissions, usually with no central view. When an auditor asks which agent can reach the HR records, the honest answer in that situation is: we will look into it.
Forty connectors, or thirteen
On the left every agent wired separately to every system. On the right the same agents and the same systems, all plugged into a single standard.
Connecting, in four steps
The order matters. Start at step three and you are granting permissions on data nobody has yet established to be correct.
Decide what the agent actually needs to touch
Not everything it could reach, but everything it needs for the task you are giving it. Every connector you do not build is a connector that cannot break and cannot leak. This step is a pruning exercise, not an inventory.
Check what already exists per system
For a lot of common software a ready-made MCP server already exists. For your own or older systems you build one, usually on top of the API that is already there. Systems with no open interface at all are the real work, and you want to know about those up front rather than afterwards.
Give every agent its own identity
An agent is a user without a face. Like any member of staff it should have its own account, its own permissions, someone who issues them and someone who revokes them. Shared keys serving several agents are exactly where an audit grinds to a halt.
Record what happens
Which agent consulted or changed which system, and when. Without that trail you cannot explain afterwards what took place, and that is precisely what a regulator asks. This step costs little if you do it up front and a great deal if you skip it.
What MCP is not
A new standard always attracts more promise than it delivers. Three things MCP does not solve for you.
Not a replacement for your systems
MCP sits alongside your software rather than in place of it. Your finance system remains your finance system. What changes is how an agent gains access to it, not what is in it or who runs it.
Not a security product
The standard governs how an agent talks to a system, not whether it is allowed to. Permissions, identity and oversight are separate work. Anyone who thinks MCP has sorted out their security has only reduced the number of doors.
No guarantee the answer is right
MCP makes sure an agent can reach your data. It says nothing about whether that data is correct or contradicts itself. Perfect connectors into messy knowledge simply produce a wrong answer more quickly. Underneath them belongs a knowledge layer.
In short: MCP is the connection, not the content. It decides whether your agents can do something rather than merely talk, and it saves you a mountain of maintenance. What comes out depends on what sits underneath, and that starts with a knowledge layer.
What people ask about this
MCP stands for Model Context Protocol. It is an open standard that lets AI models and agents connect to external systems in a uniform way: databases, documents, APIs and tools. Anthropic published it as an open standard in November 2024, after which it was adopted by several providers. The closest comparison is USB: not a different plug for every device, but one socket everything fits.
Because it determines whether your AI can do something or only talk. Without a standard you build a connector per combination of agent and system. With five agents and eight systems that is forty connectors you have to maintain yourself. With MCP it is thirteen: five agents that speak the standard and eight systems that offer it. That difference is not technical but financial, because every bespoke connector is maintenance that keeps coming back.
No, the effect is the opposite. MCP is an open standard, not a product. Because your systems offer the standard rather than one specific model, you can switch models without rebuilding your connectors. That is exactly the dependency worth avoiding: not on a protocol, but on a vendor that is the only one able to reach your data.
MCP is a connection standard, not a security product. It governs how an agent talks to a system, not whether it is allowed to. That question you have to answer separately: which agent gets which permissions, who issues them, who revokes them and what gets recorded. An agent is a user without a face, and like any member of staff it should have its own identity and its own permission profile.
MCP makes sure an agent can reach your data. It says nothing about whether that data is correct, current or self-contradictory. An agent with perfect connectors into messy knowledge simply produces a wrong answer more quickly. That is why a knowledge layer belongs underneath the connections: one place recording what is true in your organisation.
No. MCP sits alongside your systems rather than in place of them. For a lot of common software a ready-made MCP server already exists. For your own or older systems you build one, usually on top of the API that is already there. The heavy work sits with systems that have no open interface at all, and that is precisely what to examine up front rather than afterwards.
Can your AI reach your own systems?
The AI Navigator™ maps where your organisation stands, including which of your systems will connect and which will not.