In this session, Kent Dodds provides a thorough overview of MCP (Model Context Protocol) and how it is changing the way we interact with software.
You can watch the session on YouTube or read below for some selected quotes.
The Jarvis Vision
Kent opened his presentation by referencing a scene from Iron Man where Tony Stark uses his assistant Jarvis to investigate a crime scene. In just one minute of footage, Jarvis compiles databases from multiple sources, generates holographic UI on demand, accesses and relates records from different systems, and even creates a flight plan.
"My question is, what can we not do that Jarvis did in this clip? What does Jarvis do that we are not able to do? And I would say nothing with like a pretty strong asterisk."
Why Current AI Assistants Fall Short
"What's really holding us back from Jarvis is integrations. Every single one of the things that we could imagine we want to do is just really difficult to build an integration for."
"Google can't build every possible integration in the world, and those individuals can't build an integration with every other piece."
"My city is not going to build a custom GPT so I can reserve a pavilion at the park."
The Promise of MCP as a Standard
"The model context protocol says that these AI applications can integrate with these services provided that the client implements the client side of the specification and the server implements the server side. Now we can talk, and it's very similar to the way that websites work."
"The Google Chrome team doesn't have to talk to the Bob's Chocolate Factory team to be able to make that work. Bob's Chocolate Factory team makes a server that serves HTML...That's what MCP does for AI agents."
"We need a standard of communication, and that's what the Model Context Protocol is."
Designing for Agents, Not APIs
"You really should be thinking about these assistants as human, as much as possible. Like how would a human interact? That's how we want to help the LLM use our stuff."
"If you were to just give a human a page of all of your API spec and just like little inputs for all of the, you know, give me JSON, the human would probably make a mess of your database."
"I think of MCP as a replacement for a website. So like we have websites that are very, like sometimes it's just literally one thing. That's all it does. It's just a one purpose website. You can have an MCP server to do probably that exact same thing."
The Multi-Step Workflow Advantage
"What makes MCP really cool and just this workflow of working with LLMs is that it can not only just do that one task, but it also can kind of chain these tasks together."
"If you can say, okay, send money to the friend and then also schedule time on our calendar and then also send them a text to let them know I'm gonna be late or whatever. Like it all kind of, once you do all of those things altogether, then it starts to make a little bit more sense."
The Future of User Interaction
"I also think that in the next 12 to 18 months, users are just fleeing the browser as much as they can to go to these AI assistant applications. And you want to be there when they ask."
"I have used Google like maybe a 10th of the amount since ChatGPT became actually pretty good in the last year or so. I just rarely do a Google search anymore. If I need to know information, I'm going to an AI Assistant."