A language model is only as useful as the information and tools it can access. Until now, every AI application needed a separate integration for every target system, whether ticketing system, file storage, or database. The Model Context Protocol standardizes this connection and is therefore often compared to a universal plug standard. The specification was published as an open standard at the end of 2024 and is now supported by numerous AI applications and development environments.
How does MCP work?
MCP follows a client-server architecture:
- Host: the AI application a person works with, such as a chat assistant or a development environment.
- MCP client: a component in the host that maintains the connection to one server at a time.
- MCP server: a program that provides access to a specific system.
Servers offer three types of capabilities: Tools are actions the model can trigger, such as creating a ticket. Resources provide data and context, such as documents. Prompts are predefined templates for recurring tasks.
Communication uses standardized JSON-RPC messages, locally via a process's standard input and output or over the network via HTTP. The model decides which tool to use based on the task. For tool calls, the specification recommends that a person be able to confirm or reject them.
Typical examples
- An assistant reads open tickets from the service system and summarizes them.
- A programming assistant queries a project's documentation and task list.
- An internal AI agent retrieves reports from a database and creates appointments.
What companies should keep in mind
- Use only MCP servers from trustworthy sources. A server is software with access rights and can contain malicious code.
- Assign rights according to the principle of least privilege, read-only where possible.
- Plan for prompt injection: content from documents or web pages can contain hidden instructions that misuse tools.
- Require confirmation for actions that write or delete, and log them.
- Store credentials securely and authenticate remote servers properly. The specification provides an optional OAuth-based procedure for this.
How it differs
An API is the interface of a specific system. MCP is a uniform layer for AI applications and often uses exactly these APIs in the background. An MCP server thus translates between the AI application and the existing interface.
View software development