What is MCP?
MCP (Model Context Protocol) is a “common standard” that lets AI connect to external tools and data — your files, databases, the web, or the apps you use. In short, it turns AI from something that only answers from what it already knows into an assistant that can actually act on your real data.
What does MCP stand for
MCP stands for Model Context Protocol, an open standard announced by Anthropic in late 2024 so that AI (language models) can talk to external tools using one shared “common language.”
Think of MCP as the USB-C port of the AI world — previously each tool had to connect to AI its own way, but with MCP any tool that “speaks MCP” plugs into any AI that supports MCP, without writing new glue code every time.
How MCP works
MCP works on a client–server model with two main sides:
- MCP Server — wraps one tool or data source (e.g. a server for reading files, connecting to Google Drive, searching GitHub) and declares “here's what I can do.”
- MCP Client — the AI or AI-powered app (e.g. Claude, Cursor) that connects to the server to use those capabilities.
Roughly how it goes in practice:
- You connect an MCP server to the AI (a one-time setup)
- The AI sees the list of tools the server exposes, with descriptions of what each does
- When you give a task, the AI picks the right tool itself — “read this file” or “query that database”
- The server does the work and returns the result; the AI turns it into an answer
What MCP lets AI do
MCP expands what AI can access and act on, beyond just answering from its own knowledge — for example:
- Read/write files locally or in the cloud (Drive, Notion, Google Docs)
- Query and edit databases (Postgres, SQLite) without copying data out
- Pull live data from the web / APIs (prices, weather, news)
- Connect to dev tools like GitHub, Sentry, Docker
- Control apps — send a Slack message, create an issue, book a calendar slot
MCP changes what AI can access, while AI Skills add specialized know-how and AI Rules set the standards the AI must follow — three different jobs that complement each other.
Real-world MCP examples
Popular MCP servers people actually use:
- Filesystem — lets AI read/write files in a chosen folder
- GitHub — search code, read issues/PRs, create commits via AI
- Google Drive / Notion — let AI search and summarize your documents
- Postgres — ask your database in plain language and get SQL + results
- Brave Search / Fetch — let AI search the web and read pages live
See curated MCP servers, with badges for which are official / which AIs they support, in the MCP library.
Does it work with Claude, Cursor and Codex
Yes — MCP is an open standard, and major AI tools support it more and more:
- Claude (Claude Desktop, Claude Code) — full MCP support; the creator of the standard
- Cursor — add MCP servers in settings; great for coding tasks
- OpenAI Codex / Agents and others like Windsurf, Cline — increasingly support MCP
How to start using MCP
Getting started with MCP takes 3 steps:
- Pick an MCP server that fits the job — want AI to read files? Choose Filesystem. Open the MCP library and browse by task.
- Add the config into your AI tool — copy the config (JSON) into Claude/Cursor's settings file, following each tool's guide.
- Restart and try a task — reopen the AI, test a task that needs that tool; if set up right, the AI calls it on its own.
Don't want to wire them one by one? Pick a Loadout with MCP — bundled by task, ready to use.
Is MCP safe
MCP is designed so you control the permissions — AI can access only what the server exposes and you authorize. Still, keep these in mind:
- Install only trustworthy servers — pick official or well-reviewed open-source ones you can inspect
- Grant only what's needed — e.g. allow reading just the folder required, not the whole machine
- Watch actions with real impact — deleting files, sending messages; require confirmation first
MCP vs Skill vs Agent
These three are often confused, but they do different jobs:
| Difference | MCP | Skill | Agent |
|---|---|---|---|
| What it changes | What AI can access / act on | Specialized know-how | A helper that works on its own |
| Like a… | Plug for tools | Manual / built-in skill | Specialist teammate |
| When it runs | When AI needs data/tools | When a matching task appears | When the main AI delegates |
Read on: What are AI Skills · What are AI Agents · What are AI Rules
Loadouts with MCP
In practice you rarely use MCP alone — you combine several with Skills/Rules/Prompts to finish one job. A bundle assembled like that is called a Loadout.
For example, a web-dev Loadout might include MCP for GitHub + Filesystem, a code-review Skill, and a coding-standards Rule in one set — ready to use without assembling it yourself.