Skip to main content

MCP (Model Context Protocol)

MCP is like USB for AI. Before USB, every peripheral needed its own cable and driver — printers, keyboards, cameras all had proprietary connectors. USB created a single standard, and suddenly everything just worked. MCP does the same for AI applications: it provides a single, open protocol that lets any AI application connect to any tool or data source. Instead of building custom integrations for each service, you build one MCP server and every MCP-compatible client can use it.

Prerequisites​

Before reading this page, make sure you're familiar with:

  • Large Language Models (LLMs) — the AI models that MCP connects to external capabilities
  • Tool Use — how models call functions, which MCP standardizes
  • Agents — autonomous systems that benefit most from MCP's standardized connections

How It Works​

As the diagram shows, MCP uses a three-layer architecture: Host, Client, and Server.

The Host is the AI application you interact with — Claude Desktop, an IDE extension, or any tool that wants to use external capabilities. Inside the host lives an MCP Client, which handles the protocol details and maintains a 1:1 connection with a server. You don't build the client yourself; it comes with the host application.

The MCP Server is a lightweight program that exposes three types of capabilities:

  • Tools — functions the model can call (like searching a database or creating a file)
  • Resources — data the model can read (like configuration files or database schemas)
  • Prompts — reusable prompt templates that guide the model for specific tasks

Communication between client and server uses JSON-RPC 2.0 over stdio or HTTP. The protocol includes capability discovery, so when a host connects to a server, it automatically learns what the server can do — which tools are available, what resources can be read, and what prompts are offered. No manual configuration needed.

A host can connect to multiple servers simultaneously. Your Claude Desktop might connect to a GitHub server, a database server, and a filesystem server — all at once, all through the same protocol.

Why It Matters​

Before MCP, every AI tool integration required custom API code. Want your AI to search Jira? Write a Jira integration. Want it to query your database? Write a database integration. Every new data source meant new code, new auth handling, new error patterns.

MCP standardizes all of this: build a server once, and any MCP-compatible host can connect to it. This is how the ecosystem scales — developers share tools and capabilities without being locked to a vendor. Anthropic created MCP as an open standard, and it's supported by Claude, various IDE extensions, and a growing community of open-source servers covering everything from GitHub to Slack to PostgreSQL.

For teams, MCP means you can build internal servers that expose your company's tools and data, and every AI application your team uses can access them through a single protocol.