Model Context Protocol

Understand it deeply

What is MCP, really?

An open protocol that lets AI applications consistently discover, connect to and use external tools and data.

MCP does not make a model intrinsically smarter. It solves a connection problem: a chat client, IDE or agent does not need a separate bespoke integration for every document store, database or business system.

Think of a universal socket for AI

Without a shared interface, every AI application has to adapt separately to GitHub, file systems, design tools and internal services. MCP acts like a universal socket: AI applications connect through shared rules, and tool services describe their capabilities through the same rules. Server-side permissions and user approval still determine what is allowed.

How one tool call happens

1. A client connects to a server

   An AI client or agent connects to an MCP server. The server declares the tools, readable resources or prompt templates it offers.

2. The model selects a tool

   For a user task, the client presents available capabilities to the model. The model decides whether a tool is needed and supplies structured arguments.

3. The server returns a controlled result

   The MCP server performs the request within its own permission boundaries and returns a result. The model can then answer, plan another step or ask for confirmation.

The three MCP roles

MCP defines how parties cooperate; it is not itself a tool. Separating the roles helps locate whether a problem belongs to the model, client or connected service.

• 01 — MCP client

  The application hosting the AI interaction, such as a chat client, IDE or agent runtime. It connects to servers and coordinates calls.

• 02 — MCP server

  A service that exposes an external capability through the MCP specification, such as a repository, documentation system, database or internal connector.

• 03 — Tools and resources

  Tools perform actions and resources provide readable context. The protocol can describe them, but it does not bypass permissions, approvals or security policy.

What does this look like in AI coding?

A developer may ask an AI to investigate a production issue. Through MCP, the client can connect to repository, ticket and documentation services. The model reads relevant logs and rules before suggesting a fix; writes, deployments and destructive actions should still be governed by explicit permissions and confirmation.

Common misunderstandings

Is MCP the same thing as an API?

No. An API is a general software interface. MCP is a protocol for connecting AI tools and context. An MCP server will often use ordinary APIs internally.

Does adding MCP let a model operate my systems freely?

It should not. MCP standardises the connection; what can be read or done is restricted by the server implementation, credential permissions, user authorisation and product confirmation flows.

MCP is a shared way for AI applications to connect to external tools and data; it standardises access without replacing permissions, security or human approval.