What is the Model Context Protocol (MCP)?

Model Context Protocol: definition

The Model Context Protocol, MCP for short, was introduced by Anthropic in late 2024 and is openly documented. It solves a practical problem: without a standard, every AI application has to connect to every system individually. With MCP, you build the connection once as an MCP server and use it in all applications that support the protocol.

MCP is often compared to a USB port for AI: one standard plug between an AI application and systems such as file storage, database, ticket system or ERP. Many providers and tools now support the protocol, so the connections are not tied to a single model.

How MCP works

MCP distinguishes between client and server. The AI application, such as an assistant or a development environment, contains a client. The MCP server exposes what it can do: tools that carry out actions, resources with readable data and predefined prompts. Communication runs through defined messages, locally or over the network.

At startup, the client asks which capabilities a server offers and makes them available to the model, which uses them through function calling. Function calling is therefore the model’s capability, and MCP the standard for how the tools are provided. This lets the same assistant use new systems as soon as a suitable server is available.

Real-world example: an assistant for order data

A machinery manufacturer wants sales and service to be able to ask an AI assistant about orders, documents and tickets. One MCP server is created each for ERP, document storage and ticket system, offering read-only functions. The assistant can then combine order status, drawings and open tickets.

Later another assistant is added, for example for management. It uses the same servers without the connections having to be rebuilt. Access rights are controlled via the login of the respective person, so nobody sees more than before. Anyone asking new questions needs no interface programming, at most an additional server.

Security and what companies should watch for

MCP servers have access to your systems and are therefore a security matter. Use only verified servers, assign minimal rights, begin with read-only access and log calls. Content that a server returns can also contain hidden instructions, so prompt injection must be taken into account.

Nano Concepts develops MCP servers for your systems and checks rights, login and logging before an AI assistant works with them. First we clarify together which use case has the greatest benefit and start with read-only access, which we expand step by step.

• Start with one use case and read-only access
• Use only trustworthy, verified MCP servers
• Take the login and rights of the respective users into account
• Log calls and require approvals for write actions
• Plan maintenance and updates for the servers