What is an MCP server?
An MCP server is a process that exposes tools, resources, and prompts to an LLM client over the Model Context Protocol, Anthropic's open JSON-RPC standard. The client (Claude Desktop, Cursor, a custom agent) connects to the server, lists its tools at runtime, and calls them as the model decides. One server, many clients.
What is an MCP server?
An MCP server is the tool-and-data side of a Model Context Protocol connection. It speaks JSON-RPC to a client (Claude Desktop, Cursor, Zed) and exposes three primitives: tools the model can call, resources it can read as context, and prompts it can pull as templates. The official filesystem server exposes read_file, list_directory, and create_directory.
How it works
Servers run as a separate process and talk to a client over one of two transports. stdio is the default: the client spawns the server as a subprocess and exchanges JSON-RPC on stdin/stdout. SSE / streamable HTTP is for remote servers.
At connect, the client calls initialize, then tools/list, then tools/call per invocation. The model sees tool names and JSON-Schema arguments, not the protocol. MCP solves the NxM problem: one GitHub server works in Claude Desktop, Cursor, and any other MCP-aware client without per-client glue.
Where it breaks
Most engineers meet MCP when wiring an agent to internal systems. The official Postgres, GitHub, Slack, and Puppeteer servers are common starts. Gotchas: stdio servers can't hot-reload, approvals are coarse (no read-only flag), and Claude Desktop's per-chat allow doesn't persist.
For non-technical users
Think of an MCP server as a plug. Your AI assistant is the device, the server is the adapter that lets it reach one specific tool. Install once, the assistant uses it whenever it needs that tool.
Last updated: May 20, 2026