What is Model Context Protocol (MCP)?
After trying to wrap my head around MCPs for the past couple weeks, I finally get it now.
It helps me to think about it backwards:
Protocol → Context → Model
MCP is a protocol (a standard way of doing something) for providing context (structured, relevant information) to a model (like Claude 4 Opus, Gemini 2.5 Pro, or whatever LLM you’re working with).
That's the gist of it.
Okay, but what is it?
Last week at the first Cursor meetup in Toronto at 535, I had the opportunity to ask Dan Sadler, CTO of Rootly, about MCPs. He summed it up nicely:
“MCP is an abstraction on top of APIs to talk to them with natural language.”
Where did MCPs come from?
Mike Krieger is Chief Product Officer at Anthropic, the company that created several jaw dropping LLMs, including Claude 3.5 and the newly announced Claude 4 Sonnet and Opus. MCPs were born at Anthropic. Mike shared the origin story during a Sequoia talk titled “Building AI Products From the Bottom Up.” https://www.youtube.com/watch?v=Js1gU6L1Zi8
The Anthropic team built a "connector" for Google Drive into Claude to give it access to a user's docs, sheets and other files. Those files are useful "context" if you want your "model" to help you with work.
Next, they built a "connector" for GitHub to give the models access to a repo's source code, PRs, git commit messages, and more.
These two connectors were built completely differently.
As a third "connector" was being scoped out, someone asked: shouldn't these connectors have more in common? MCP came out of that moment. It was two engineers realizing there was a pattern here.
Hence, a protocol for providing context to models was born. That's mcp.
P.S. If you're wondering why Mike Krieger sounds so familiar. He co-founded Instagram with Kevin Systrom.
Why does this matter?
Because context is what makes AI models really useful.
“Context” means everything from “read this document” to “fetch the docs for the latest version of this library” to “mark this ticket as done in Linear.” Without a shared structure for these interactions, every integration becomes bespoke and brittle.
MCP fixes this.
Want to browse browse through literally thousands of MCPs? Visit smithery.ai.
So, what's next?
MCPs are heading in two major directions:
-
Taking action
The first wave of MCPs focused on pulling in information. But going forward, enabling models to take actions. This is how you move from just chatting with models to actually getting things done. Remember Dan Sadlers comment about MCPs being "an abstraction for APIs." Pretty sure there'll be MCPs for things we currently do with APIs, and more. -
Agent-to-Agent collaboration
As agents become more capable, the question becomes how they talk to each other. It’s an exciting space. We'll see agents delegating (aka, hiring) tasks to other agents. Seamless communication between agents will be critical for this. Agent2Agent (A2A) by Google is an open standard that's catching on that aims to facilitate this. But that's a story for another day blog post.
Let's Recap
-
MCP = Model Context Protocol
-
It’s a protocol for providing context to models.
-
MCPs help models not just retrieve context, but also take actions and potentially collaborate with other agents.