What it is. MCP is Anthropic's open protocol (introduced in late 2024) for connecting AI assistants to external tools and data. An MCP server is a small program — usually running locally — that speaks MCP and exposes capabilities: tools (function calls), resources (readable files), or prompts (reusable templates). Any MCP-compatible client — Claude Desktop, Claude Code, Codex, Cursor, Cline, self-hosted agents — can then use it without knowing how it works internally.
Why it matters. Before MCP, every AI-tool integration was custom glue code. MCP standardizes the interface, so an MCP server written once works everywhere. Cursor calling the same GitHub MCP server as Claude Desktop, without either side knowing about the other, is the point of the protocol.
What to check before adopting. An MCP server runs with whatever permissions you hand the process — file access, shell execution, tokens, network. That's why every MCP server on AgentMaps carries a risk level and an explicit permission scope. Treat installing an MCP server the same way you'd treat installing any other program: read what it can do before you wire it into a real project.