Architecture

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that defines how AI models and agents connect to external tools, data sources and business systems, so that a single connection can be reused by any compatible AI application.
DATE UPDATED: 09 Oct 2026
|
FIRST PUBLISHED: 09 Oct 2026

What it means in practice

Before MCP, every AI tool needed its own bespoke integration with every system it touched. Connecting an assistant to a CRM, a document store and a finance platform meant three separate pieces of work, and switching to a different AI provider meant doing them again. MCP replaces that with a common standard. A system is exposed once through an MCP server, and any AI application that speaks the protocol can discover what it offers and use it.

Anthropic introduced MCP in late 2024 and donated it to the Agentic AI Foundation, under the Linux Foundation, in December 2025. It’s now supported across the major AI platforms, which is what turned it from an interesting idea into infrastructure.

For a business, the practical effect is that connecting existing systems to AI becomes considerably cheaper and less locked to any one vendor. The harder questions move elsewhere: which systems to expose, what an agent is allowed to read or change through each connection, and how that access is logged and revoked.

How it differs from an API and from function calling

An API is the interface a specific system offers to other software, and every system’s API is different. Function calling (or tool use) is the mechanism a model uses to invoke a tool mid-task. MCP sits between the two. It standardises how tools and data are described, discovered and connected, so a model doesn’t need custom code to understand each system’s API before it can call it.

Insight

For organisations with years of existing infrastructure, MCP shifts where the difficult work sits: the plumbing gets easier, and the design decisions around it carry more weight. When connecting a system to an agent takes days rather than months, the temptation is to connect everything and work out the rules later, which is how firms end up with agents holding wider access than any individual employee would be given.

The decision about what to expose, to which agents and with what permissions is an architectural one, and it belongs before tool selection rather than after it. MCP makes that decision easier to implement, though someone still has to make it.

Misconceptions

MCP is a protocol that products implement (much as email works across providers), so there's nothing to buy as such.

A connection made through MCP isn't safe by default either. The standard defines how systems talk, and security depends on how each server is built and permissioned. Over-broad access and instructions hidden in tool outputs are both live risks.

It doesn't fix the data underneath. An agent connected cleanly to a messy, inconsistent system will produce messy, inconsistent results faster.

BACK TO GLOSSARY