MCP (Model Context Protocol)
MCP is an open standard that lets AI models access external tools, data, and systems in a uniform way, instead of building a custom integration for every connection.
MCP stands for Model Context Protocol. The open standard describes how AI applications can connect external tools, data sources and workflows in a consistent way.
An application can use standardised MCP servers instead of defining a completely separate integration for every external system. Actual compatibility depends on the supported capabilities and permissions.
How MCP is structured
MCP distinguishes between host, client and server. The host application coordinates the AI functionality and manages a separate client connection for each connected server.
An MCP server provides bounded functionality and context. This may include executable tools, readable resources and reusable prompt templates.
The server describes its supported capabilities in a standardised format and may run locally or as a remote service.
An MCP client is a component of the host application and maintains one connection to an MCP server. It handles capability negotiation and the exchange of protocol messages.
The application decides how the available tools, resources or prompts are incorporated into the interaction with the model.
Why MCP matters for agentic systems
Without a shared protocol, different AI applications often require separate integrations for the same internal systems. This increases engineering, maintenance and review effort.
An MCP server for an internal system can be used by several compatible host applications. This supports reuse of a reviewed integration, for example for a CRM or knowledge base.
The actual switching effort still depends on permissions, supported capabilities and provider-specific logic outside MCP.
Security questions that MCP raises
MCP servers can provide access to production data and actions. The host application and server must therefore enforce permissions, user consent and security boundaries consistently.
Define which actions are allowed, which require confirmation and how calls are logged. Write access in particular should be limited according to the principle of least privilege.
Further reading
Related terms
Related topics
Related services
Would you like to apply this in your organisation?
Discuss your project