What is MCP?
📖 5 min readUpdated 2026-04-18
MCP (Model Context Protocol) is an open standard, introduced by Anthropic in late 2024, for connecting tools and data to AI models. Think of it as the USB-C port of AI.
The problem MCP solves
Before MCP, every time you wanted to give an AI app access to a new tool. Gmail, GitHub, your local file system, someone had to write custom glue code. That glue was non-portable: it worked with one app and one model, and had to be rewritten for the next.
MCP standardizes the glue. Write one MCP server for Gmail and any MCP-speaking client (Claude Code, Claude.ai, other agents) can use it. Build once, plug in anywhere.
The one-sentence pitch
"A standard way for AI agents to discover and use external tools and data."
What MCP actually does
An MCP server exposes three things to an MCP client:
- Tools, functions the model can call (e.g.
search_email, create_issue)
- Resources, data the model can read (e.g, documents, database records)
- Prompts, reusable prompt templates
The client (e.g. Claude Code) connects to the server, asks what's available, and makes those tools/resources/prompts available to the model.
Why it matters for autonomy
Autonomous agents need many tools to do real work. Without MCP, connecting an agent to 20 services means writing 20 custom integrations and maintaining them as APIs change. With MCP, you install 20 pre-built servers (your own or from the community) and the agent uses them immediately.
Insight: MCP doesn't make a model smarter. It multiplies what the model can reach. That's what makes autonomous work viable.
Who uses MCP today
- Claude Code. Anthropic's CLI agent, MCP-first
- Claude.ai, the web app supports remote MCP servers for "connectors"
- IDEs. Cursor, Zed, and others speak MCP for in-editor tools
- Custom agent frameworks. LangGraph, CrewAI, and others have MCP adapters
Who builds MCP servers
- Anthropic, reference servers for common integrations
- Service providers, companies (Notion, Slack, Figma, Canva) ship official MCPs so agents can use their products
- The community, hundreds of open-source MCPs on GitHub for everything from databases to home automation
- You, writing a custom MCP for your own tool is a few dozen lines of code
What MCP is NOT
- Not a model. MCP is a protocol. The model (Claude, GPT, whatever) still does the thinking.
- Not tied to one vendor. It's an open spec. GPT and Gemini can (and will) speak MCP too, even if the early momentum was Anthropic-led.
- Not magic. A bad tool is still bad, wrapped in MCP or not. Design discipline still applies.
What to read next
Head to Anatomy to understand the five primitives, or jump to Directory for a curated list of MCPs worth trying.