What is MCP (Model Context Protocol)? A plain-English explanation
If you've seen "MCP" all over your feed and quietly wondered what it actually is, this is for you. No jargon wall, promise.
MCP (Model Context Protocol) is an open standard that lets AI apps use tools and data through one common interface. Build one MCP server for your system, and any AI app that speaks MCP can use it. That's the whole pitch.
Everything below comes from the official spec. At the time of writing (19 October 2026) the current revision is 2026-07-28.
Why anyone bothered inventing it
Say you have five AI apps and eight systems they should reach: a database, a CRM, a docs store, a ticket tool, and so on. Without a standard, that's up to 40 separate integrations. Somebody has to build and maintain every one of them. Nobody wants that job.
With a standard, each system gets one MCP server. Each AI app learns one way to talk to servers. Done.
Three roles you need to know
- Host: the AI app a person actually uses. It owns the model, the screen and the "are you sure?" prompts.
- Client: the connector living inside the host. One client per server.
- Server: the program that offers capabilities. It can run on your laptop (the host starts it) or live online as a service.
A local server talks over stdio. An online one talks over Streamable HTTP. That's it, those are the two standard transports.
What a server can offer
| Thing | What it is | Example |
|---|---|---|
| Tools | Functions the model can ask the host to run | search_docs, gst_breakdown |
| Resources | Read-only data with an address (a URI) | docs://catalog |
| Prompts | Ready-made prompt templates a user can pick | "Summarise this ticket" |
Clients can offer a few things back too (asking the user for input, for example), but those are optional.
How the messages look
Under the hood it's JSON-RPC 2.0. In the current revision there is no handshake step any more. Every request carries its own version and capabilities. We wrote a separate post on that, with a real captured request: what changed in MCP 2026-07-28.
When you should NOT use MCP
Honestly, a lot of the time. If one web app calls one internal service, and nobody else ever will, a plain REST call is easier, cheaper and simpler to secure. MCP starts to pay off when several AI apps, or your users' own AI tools, need the same capability. We made a proper decision table here: MCP vs REST vs function calling.
What MCP will not do for you
- It won't make a model trustworthy. If a tool returns text with instructions hidden inside, a model may follow them. That's called prompt injection. Our MCP security checklist covers it.
- It won't log your users in. Authorization is optional in the spec, so you have to add it.
- It won't check your business rules. Your server code still has to say no to bad input.
Where to go next
- Run a real server today with the Python MCP server tutorial.
- Keep the MCP cheat sheet open in a tab.
- Want the whole path in one place, with diagrams and tested code? That's the MCP Builder's Handbook.
Frequently asked questions
What does MCP stand for?
MCP stands for Model Context Protocol. It's an open protocol that lets AI apps talk to tools and data in one standard way.
Is MCP an API?
Not really. An API is the door to one particular service. MCP is the shared language an AI app uses to find and call tools on many different servers. Often an MCP server just wraps an API you already have.
Do I need MCP to build AI features?
No. If only your own code calls the thing, a normal API call or plain function calling is usually simpler.