Concept guide
MCP explained: what the Model Context Protocol is and why it matters
What the Model Context Protocol (MCP) is, how clients, servers, tools and resources fit together, a simple example, the security risks, and when to use it.
Short answer: the Model Context Protocol (MCP) is an open standard that lets AI applications connect to tools and data in one common way. Instead of writing a custom integration for every tool and every assistant, you build an MCP server once and any MCP-capable application can use it.
The problem it solves
Without a standard, connecting five assistants to ten tools means up to fifty custom connectors. MCP turns that into one connector per tool and one client per application, in the same spirit as a USB port for devices.
The pieces
- Host: the AI application the user talks to (MCP host).
- Client: the part inside the host that keeps a connection to one server (MCP client).
- Server: a small program that exposes capabilities (MCP server).
- Tools: actions the model may ask to run, such as "search tickets" (MCP tool).
- Resources: read-only data the application can load, such as a file or a record (MCP resource).
- Prompts: reusable prompt templates a server can offer.
A simple example
A company runs an MCP server for its helpdesk. It exposes a tool search_tickets(query) and a resource for each
knowledge-base article. An assistant connects, discovers the tool and its description, and when a user asks about a
refund it calls search_tickets("refund"), reads the result and answers. The same server can later work with a
different assistant without changes.
Security: what to watch
An MCP server can act with real permissions, so treat it like any service account:
- Give each server the narrowest access that works (read-only where possible).
- Treat tool descriptions and tool results as untrusted text: they can contain prompt injection.
- Require human approval for actions that send, pay, delete or publish.
- Log every tool call. See guardrails.
When to use MCP
Use it when several applications need the same tools, or you want tools to be swappable. For a single app with one or two tools, plain function calling is simpler. To build one, follow build an MCP server in Python.
Where it fits
MCP is one layer of agentic systems: the model decides, the client carries the request, the server does the work. Glossary: Model Context Protocol. Official documentation: modelcontextprotocol.io.