MCP (Model Context Protocol) is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools. - modelcontextprotocol.io

Honestly, I can’t remember the last time I was this excited about a new technology—let alone a protocol. I’m more excited about MCP than I was when I first tried GPT-3.5.

Why? Because at their core, LLMs are glorified document completers. They’re trained on massive datasets from across the internet and designed to behave like conversational agents. You can ask them questions, and they’ll often give you thoughtful, accurate responses. But they have two fundamental limitations:

  1. They can’t access or reason over private data.
  2. They can’t take actions on your behalf.

An LLM that can take action is at least 10x more useful than one that can just answer questions. If there’s a task to be done, such an LLM could automate it entirely—removing the need for you to even be in the loop. This is the promise of MCP. And we’re only just beginning to scratch the surface.

Here’s a great example of a hypothetical agent that uses MCP to take action on your behalf:

mcp-overview

The most powerful thing about MCP is that it gives you instant access to thousands of pre-built MCP servers, which you can use to build your own custom agent. In the example above, you might find MCP servers already built for Doordash, Google Calendar, and Twilio—meaning you don’t need to worry about building those API integrations yourself. Your LLM will know how to communicate with these services via their respective MCP servers.

Even better: there’s already a growing list of compatible servers, and the team at Anthropic is working on a discoverable registry of MCP servers — launching very soon.

Further Reading

Want to get a deeper dive into how MCP works.