What is MCP? The Model Context Protocol explained
MCP is the USB-C of AI tooling: a single standard that lets any model talk to any tool. Here's what it is, why it matters, and how to get started in 10 minutes.
MCP (Model Context Protocol) is an open standard, created by Anthropic in late 2024, that defines how AI models communicate with external tools and data sources. In plain terms: instead of manually copying context into a chat window, MCP lets Claude (or any MCP-compatible model) call tools directly — read a file, query a database, run a browser — and get structured results back. It's the difference between an AI advisor and an AI that can actually do things.
The analogy that sticks: MCP is the USB-C of AI tooling. Before USB-C, every device had its own proprietary connector. USB-C created one standard that works everywhere. MCP does the same for AI tools — one integration protocol, usable across Claude, GPT-4, Gemini, and any model that adopts the spec. A server you build once works with every MCP client.
How MCP works, without the jargon
An MCP setup has two parts: a client (the AI model) and a server (the thing that exposes tools). The server declares what it can do — "I can read files", "I can run SQL queries", "I can fetch web pages" — and the client can call those tools during a conversation. The AI decides when to use them based on context.
Concretely: you're talking to Claude in Claude Desktop and you ask "how many users signed up this week?". Without MCP, Claude can only answer if you paste in the data. With a Supabase MCP server configured, Claude calls the query tool, gets the number from your database, and answers directly. The conversation never left Claude's interface.
MCP vs function calling: what's the difference?
If you've used OpenAI's function calling or Claude's tool use, MCP will feel familiar — but there's a key architectural difference. Function calling defines tools inline in the API request (each call, each time). MCP is a persistent, separate process: you run a server, the client connects to it, and the connection stays open across the whole conversation. This means MCP servers can maintain state, hold open connections, and be shared across multiple clients and models.
Another difference: MCP is local-first. Most MCP servers run on your machine as a subprocess, which means your data never leaves your environment unless the tool explicitly calls an external service. This is important for developers who need to keep code or data local.
Getting started: 10 minutes to your first server
The fastest path to MCP is Claude Desktop + the official filesystem server. Install Claude Desktop, then add this to your config file (~/Library/Application Support/Claude/claude_desktop_config.json on Mac):
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/yourname/projects"
]
}
}
}Restart Claude Desktop and the model now has read/write access to your projects folder. Ask it to "list all TypeScript files in my current project" or "read the README and summarize the setup steps" — it'll call the filesystem tools and answer from the actual files.
Security model
MCP servers only have access to what you explicitly grant. The filesystem server only reads/writes paths you specify in the config. No tool can access anything outside its declared scope — the model cannot "jailbreak" a server to access your whole machine.
The MCP ecosystem in 2026
When Anthropic open-sourced MCP in November 2024, the ecosystem was a handful of official servers. By mid-2026, there are hundreds: Slack, Linear, Notion, GitHub, AWS, Google Drive, Postgres, MongoDB, Stripe, and more. The community has built servers for almost every developer tool and SaaS product.
Adoption has spread beyond Claude too. OpenAI added MCP support to their agents framework. Google's Gemini supports MCP in Vertex AI. The standard is becoming the default integration layer for agentic AI across the industry.
MCP servers to start with
MCP is the biggest practical upgrade to daily AI use since streaming responses. Once you have one server running, the question of "what else can I connect?" becomes an interesting one. The full catalog of community MCP servers is on the Radar.
Find these on the Radar
Every tool here lives on Kapyn Radar. Save the ones that fit into a Loadout and find them again.
