Model Context Protocol

The problem is combinatorial: M AI applications each integrating with N systems is M×N bespoke integrations. MCP makes it M+N by defining a common interface — a server exposes tools, resources and prompts; any compliant client can use them.

Architecturally this puts an integration boundary where there was previously bespoke glue, which means the usual boundary concerns apply and are frequently neglected: an MCP server is a privileged component that an agent will call with attacker-influenced arguments, so it needs its own authorisation, its own input validation, and its own audit log.

The strategic value is that tool integrations stop being tied to one AI vendor's SDK, which is the same reason open protocols usually win.