The "consumer side" of the MCP architecture — Claude itself, along with <a href="/en/glossary/claude-tools/claude-desktop/">Claude Desktop</a> and <a href="/en/glossary/claude-tools/claude-code/">Claude Code</a>, are all MCP Clients. They actively call tools provided by MCP Servers to complete tasks assigned by users.

An open standard launched by <a href="/en/glossary/core-concepts/anthropic/">Anthropic</a> letting AI models connect to external tools and data sources in a unified way. Before MCP, every AI application needed to write its own integration code for different tools; with MCP, developers only need to build one "<a href="/en/glossary/mcp-tools/mcp-server/">MCP Server</a>" and any MCP-supporting AI client (like <a href="/en/glossary/claude-tools/claude-desktop/">Claude Desktop</a>) can use it directly. Analogy: MCP is like USB for the AI world — same interface, any device can plug in.

The service-side component in the <a href="/en/glossary/mcp-tools/mcp-protocol/">MCP (Model Context Protocol)</a> ecosystem responsible for "providing tools and resources." Developers implement in Python or TypeScript, defining three primitive types: Resources (readable data), Tools (executable functions), and Prompts (preset prompt templates), letting MCP Clients (like <a href="/en/glossary/claude-tools/claude-desktop/">Claude Desktop</a>) call these capabilities. Building custom MCP Servers is the core technology for connecting Claude to any external system.

MCP Transport is the spec defining how Claude and an <a href="/en/glossary/mcp-tools/mcp-server/">MCP Server</a> exchange messages. The two main options are stdio (standard input/output) and SSE/HTTP. stdio suits local connections on the same machine and is the simplest to configure; SSE/HTTP goes over the network and suits cloud deployments and team use. Which you pick depends on whether your server and Claude run on the same machine.