Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open specification that describes a single, vendor-neutral way for AI applications to feed an LLM the context it needs — the tools it can call, the data it can read, and the prompts it can reuse. Rather than hand-coding a separate connector for every model-and-service combination, you implement an integration once behind the protocol, and any compatible host can speak to it. It is an emerging standard: first published by Anthropic in late 2024, adopted by a growing set of clients and tool authors since, and still actively evolving, so details may shift over time.

How it works: MCP follows a client–server pattern. An MCP server wraps a capability — a file system, a database, a SaaS API, a search index — and advertises it in a standard shape. The MCP client, embedded in the host app that runs the model, discovers those capabilities at runtime and relays the model's requests to them. When the agent decides to act, the request travels through ordinary tool calling, produced by the model's function calling ability, and the protocol simply standardizes the wire format so the same server is reusable everywhere.

Why it matters: integrations are usually the most brittle part of an agent. Every new tool means another bespoke adapter, and swapping models can mean rewriting all of them. By defining one contract, MCP turns tools into portable building blocks — a connector written by one team can be dropped into another team's AI agent with no rewrite, which cuts duplication and reduces lock-in to any single provider.

Concrete example: suppose your company exposes its ticketing system as an MCP server with two tools, search_tickets and create_ticket. A support agent, a coding assistant, and an internal chatbot can all connect to that one server and use the same tools — no team has to reimplement the ticketing logic, and if you change the model behind any of them, the connector keeps working untouched.