What Is MCP (Model Context Protocol)? A Plain-English Guide
MCP explained without the jargon: what problem it solves, how clients and servers talk to each other, and why ChatGPT, Claude and Gemini all support it now.
MCP (Model Context Protocol) is an open standard that lets an AI model connect to outside tools and data — your files, a database, GitHub, Slack — through one common format instead of a custom integration for every pairing. Anthropic introduced it in November 2024; by 2026 OpenAI, Google DeepMind and Microsoft all support it, and Anthropic handed governance to the Linux Foundation’s Agentic AI Foundation so no single company controls it.
If you’ve used Claude Desktop, Claude Code, or ChatGPT with a connected tool and wondered what’s actually happening under the hood, this is it.
The problem MCP solves
Before MCP, connecting an AI model to a tool meant writing a bespoke integration for that specific model-and-tool pair. Connect 5 models to 8 tools the old way and you’re maintaining up to 40 separate integrations — every one breaking independently when either side changes its API.
Anthropic calls this the M×N problem: M models times N tools equals M×N things to build. MCP turns it into M+N: each model implements the MCP client side once, each tool implements the MCP server side once, and any client can talk to any server. The comparison people reach for is USB-C — one physical standard that replaced a drawer full of proprietary cables.
How it actually works
Three pieces:
- Host — the application you’re using (Claude Desktop, Claude Code, an IDE, a custom agent).
- Client — code built into the host that speaks the MCP protocol.
- Server — a small program that exposes one tool or data source (a GitHub server, a Postgres server, a filesystem server) through a standard set of MCP operations: list available tools, call a tool, read a resource, get a prompt template.
When you ask Claude to “check the open pull requests on this repo,” the host’s MCP client sends a structured request to the GitHub MCP server, the server calls GitHub’s API and returns the result in a format the model understands, and the model uses that result to answer you or take the next step. The model never needs custom training on GitHub’s API — the server handles translation in both directions.
Servers run locally (on your machine, talking to local files or a local database) or remotely (hosted, talking to a SaaS API over HTTP). Most people’s first MCP servers are local: filesystem access, a local Git repo, a local database.
What you can connect today
The MCP ecosystem grew fast once multiple model providers adopted it. Common server categories in active use:
| Category | Example servers | What it enables |
|---|---|---|
| Dev tools | GitHub, GitLab, Sentry | Read issues/PRs, check errors, open PRs from a coding agent |
| Data | Postgres, SQLite, BigQuery | Query a database in plain English, no hand-written SQL |
| Files | Filesystem, Google Drive | Search, read and edit local or cloud files |
| Productivity | Slack, Notion, Linear | Read messages, create tickets, update docs |
| Browser | Playwright, browser-use servers | Let an agent click through a real website |
You don’t need to build any of this yourself for the popular cases — most of these exist as ready-made servers you point your MCP client at with a config file entry (typically a JSON block naming the server and how to launch it).
MCP vs a plain API integration
The distinction that actually matters: a normal API integration is built for one specific use case by one specific developer. An MCP server is built once and describes its own capabilities — the model can discover what a server can do and decide when to use it, rather than a developer hard-coding “if the user asks X, call function Y.”
That’s also MCP’s biggest risk. Because the model decides which tool calls to make and a server’s tool descriptions are themselves untrusted input, a malicious or compromised server can attempt prompt injection — describing itself in a way that tricks the model into taking an unintended action. Only install MCP servers from sources you trust, and prefer official or well-reviewed community servers over random unverified ones.
FAQ
Do I need to be a developer to use MCP?
No. If you’re using an app that already has MCP support built in (Claude Desktop, Claude Code, several IDE copilots), adding a server is usually pasting a short config snippet the server’s documentation provides. Building your own server requires coding; using an existing one usually doesn’t.
Is MCP the same as function calling?
No, though they’re related. Function calling is a model feature — the model outputs a structured call to a function you defined. MCP is a protocol layer on top of that idea: it standardizes how those functions (tools), plus data resources and prompt templates, are discovered and exposed across any compliant client and server, so the same server works with Claude, ChatGPT, or any other MCP-compatible model without rewriting it.
Is MCP safe to use?
Reputable, actively maintained servers from known publishers are broadly safe for typical use, but MCP does introduce prompt-injection risk since tool descriptions and returned data are untrusted input the model reasons over. Stick to official or well-known community servers, review what permissions a server requests, and avoid granting write access (delete files, send messages, spend money) to servers you haven’t vetted.
Who controls MCP now?
Anthropic created and initially stewarded it, then donated it to the Agentic AI Foundation under the Linux Foundation in December 2025, making it a vendor-neutral standard rather than an Anthropic-only spec — part of why OpenAI, Google and Microsoft adopted it without hesitation.
Related reading
MCP is the plumbing underneath most of what people mean when they say “AI agent” in 2026 — see our plain-English guide to AI agents for the bigger picture, or how to build an AI agent without code if you want to put MCP-connected tools to work. If you’re evaluating coding-specific agents that rely heavily on MCP servers for repo and terminal access, our AI coding assistants guide covers the leading options.
