If you shipped anything with Claude, Cursor or Windsurf in the last year, you have probably heard the acronym MCP. It is the glue that turns a language model into an actual working agent. This post explains what MCP is, why it matters, and what to install first.
MCP in one sentence
MCP — the Model Context Protocol — is an open specification from Anthropic that defines how an AI assistant connects to external tools, data and actions. Think of it as USB-C for AI agents: one cable, any device.
The problem MCP solves
Before MCP, every IDE, chat interface and assistant had its own way to expose tools. If you wanted Claude to read your Postgres database, you wrote a bespoke function. If you wanted Cursor to do the same, you wrote another bespoke function. Multiply that across GitHub, Slack, Figma, Stripe and every tool in a developer’s stack — and you have a mess.
MCP standardises the handshake. Write a server once; any MCP-aware client can use it. Write a client once; it picks up every MCP server ever built.
How MCP actually works
At a high level, three actors talk to each other:
- Host — the app the user interacts with (Claude Desktop, Cursor, Windsurf, your custom Agent SDK code).
- Client — the connection manager inside the host, speaking MCP.
- Server — the tool provider. Exposes
tools,resourcesandpromptsto the host.
The transport is usually JSON-RPC 2.0 over stdio (for local servers) or HTTP + Server-Sent Events (for remote ones). The model picks a tool, fills its parameters, the server runs it, the result flows back into the conversation.
What a typical MCP server exposes
Three primitive types:
- Tools — model-callable functions with JSON-schema arguments. Example:
search_repos(query, limit). - Resources — read-only content the model can subscribe to. Example: contents of a file.
- Prompts — reusable prompt templates the host can surface to the user.
Who is using MCP?
As of April 2026, the protocol is supported natively by:
- Claude Desktop + Claude Code
- Cursor
- Windsurf
- Zed
- ChatGPT “connectors” (via compatibility shim)
And there are 1000+ public MCP servers for everything from Postgres to Figma to web scraping.
Your first three MCP servers
If you are just starting out, install these and feel the difference:
- Filesystem MCP — read and write files in a sandboxed folder.
- GitHub MCP — pull requests, issues, code search.
- Memory MCP — persistent knowledge graph across sessions.
Why MCP matters in 2026
Two trends converge here. First, agents moved from demos to production. Second, model vendors are commoditising the raw API — the value has shifted to what an agent can do. MCP is the fabric that decides capability.
If you build products, you care because adopting MCP means you join an ecosystem of hosts and tools for free. If you build internal tools, you care because a well-scoped MCP server is safer than an ad-hoc plugin.
Where to go next
Browse the full directory of 130+ MCP servers, or jump to our step-by-step Claude Desktop install guide. If you want to build your own server, the Node.js tutorial walks through a minimal working example in 50 lines.