AIght_
ToolsLearnFieldsUniverseSignalHumanAbout
Take the quiz
← All concepts

Concept

Model Context Protocol

The open standard that lets AI models talk to your tools without a custom integration per model

Mankaran Singh·Updated May 17, 2026

Where this idea lives

PREREQUISITESTOOLS THAT SHOW ITModel Context ProtocolFunction CallingFunction Calling — The JSON-shaped API that turned chat models into clients of the real world.AI AgentsAI Agents — When AI stops answering and starts doing — and then, very often, hits a wallMultiagent SystemsMultiagent Systems — What happens when AI models try to coordinate with each other — and the new failure modes that come with itPrompt EngineeringPrompt Engineering — The craft of talking to a model that will take you exactly as literally as it decides toCursorCursorClaudeClaudeWindsurfWindsurfCommon misconception: MCP is a Claude-only thing.Common misconception: MCP lets the model do anything on your machine.Common misconception: MCP servers are hard to write.
prereqsrelatedtoolsmisconceptions
shows up in:Software Engineering
You might think:MCP is a Claude-only thing.MCP lets the model do anything on your machine.MCP servers are hard to write.

Every AI integration used to be handmade. You'd write custom code to connect your assistant to a database, a filesystem, a calendar API, a search engine. Every team building AI products was solving the same plumbing problem in a slightly different way, with slightly incompatible pipes, and none of it transferred.

Anthropic released the in late 2024 to solve this.[·] MCP is a standard for how AI models and external tools talk to each other — not a product, not a service, just a spec anyone can implement. The idea is that if everyone uses the same interface, you build the plumbing once.

Think USB. Before USB, every peripheral had its own connector, its own driver, its own installation ritual. USB said: one port, one standard. MCP is trying to be USB for AI.

Tools expose themselves as MCP servers. AI assistants connect as MCP clients. The model doesn't need to understand how the tool is built — just what it can do and how to ask.


How it actually works

An MCP server is a small program that exposes a set of tools — functions the AI can call — and optionally resources, which are pieces of data the AI can read. The server describes what it offers (name, description, input schema), and the client (an AI assistant) queries for available tools before deciding whether to use them.

The protocol runs over standard transports: standard input/output for local tools, HTTP for remote ones. That simplicity is deliberate. An MCP server can be written in any language. Existing tools can be wrapped with minimal code.

// A minimal MCP server exposing a "read_file" tool
const server = new Server(
  { name: "filesystem", version: "1.0.0" },
  { capabilities: { tools: {} } }
);

server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [{
    name: "read_file",
    description: "Read the contents of a file at a given path",
    inputSchema: {
      type: "object",
      properties: { path: { type: "string" } },
      required: ["path"],
    },
  }],
}));

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { path } = request.params.arguments as { path: string };
  const content = await fs.readFile(path, "utf-8");
  return { content: [{ type: "text", text: content }] };
});
“One AI client, any tool. Any AI client, one tool. That combination didn't exist before.

◉ INTERACTIVE

File Systemread · writeclick to connectWeb Searchquery · fetchclick to connectCalendarevents · timeclick to connectDatabasequery · storeclick to connectAI ModelMCP client

What this unlocks

Cursor is the clearest example of MCP in practice. When you ask it to read a file, edit code, or run a terminal command, it's using MCP to call standardised tools — filesystem access, code execution, search. Another AI client could connect to the same MCP servers and do the same work. The tools are no longer locked to a specific assistant.

The fragmentation in AI tooling was mostly friction, not necessity. A database tool written for Claude shouldn't need to be rewritten for GPT or Gemini. With MCP, it isn't.

This matters because every AI product was previously rebuilding the same wheel — file access, web search, database connectors. With MCP, you build it once, connect anywhere that speaks the protocol. The same MCP server runs Claude Desktop, Cursor, Zed, and anything else that wires it up.

By mid-2026, the public MCP registry lists hundreds of servers — for GitHub, Slack, Linear, your own filesystem, custom company tools.[·]


What to watch for

MCP is still young, and the ecosystem is building out quickly. A few things worth understanding before you dive in.

The quality of the server description matters a lot. The AI decides whether to use a tool — and how — based on its name and description. A well-described tool gets used correctly. A vague one gets ignored or misused in ways that are hard to debug.

If you're writing an MCP server: the tool description is the prompt. Treat it that way. A 50-word description that explains when to use this will beat a 5-word description that just names the function.

Local vs remote changes the trust model. Local MCP servers (running on your machine) have access to whatever your process has access to — files, the network, system calls. That's powerful and also worth thinking carefully about before adding untrusted servers. The ecosystem needs good security conventions, and they're still forming.

Composability is the real promise. The long-term bet is that you'll have a collection of MCP servers — one for your calendar, one for your codebase, one for your email — and any AI client you use will be able to work with all of them, without any of them knowing about each other.


Most AI products that feel genuinely integrated — not just chatting with you, but actually operating in your environment — are doing something like this. Before MCP, that meant custom code everywhere. The standard is still early, but the direction is right: the tool layer should be shareable.

← Back to all conceptsBrowse tools →
intermediate
Read time7 min read
UpdatedMay 2026
Sources2

Read next

  1. AI Agents →
  2. Multiagent Systems →
  3. Prompt Engineering →
  4. Function Calling →