What is an MCP (Model Context Protocol)?

Before MCP, every AI product that wanted to call a third-party API — search a database, read a file, hit a SaaS tool — had to write bespoke integration code for that pairing. MCP replaces that N-times-M integration problem with a single protocol: any MCP-compliant server exposes its tools, resources, and prompts in a standard shape, and any MCP-compliant host can consume them without custom glue code, the same way HTTP lets any browser talk to any web server. The protocol defines a handful of primitives — tools (callable actions), resources (readable data), prompts (reusable templates), sampling (server-initiated model calls), and elicitation (server-initiated user input requests) — carried over a transport like stdio or Streamable HTTP, with JSON-RPC as the message format underneath. Servers declare their capabilities at connection time, and hosts and clients negotiate which features both sides support. Adoption has been rapid because it turns integration work into a one-time cost: build an MCP server once for a service, and it becomes usable by any MCP-compatible AI product, not just the one it was originally built for.