There’s something I’ve been pondering lately. It’s about the concept of Model Context Protocol (MCP), and I must admit it’s really a big deal, a crucial piece of the AI and tech world that most of us are, either by choice or lack of understanding, overlooking. The reason being, the crowd seems to be narrowly focused on the enhancement of desktop applications with an “agentic” functionality—AI’s ability to perform actions in the world. But keep in mind, if you genuinely want to professionally develop agentic AI applications, you need a broader vision.
So let’s dive in and unpack this a bit more.
This video is from Confluent Developer.
Imagine this protocol as an engine under the hood of your AI vehicle. One of its fundamental roles is to transform the AI application into what we can refer to as an ‘agent.’ An agent capable of creating impacts in the ‘real’ world, just like a car engine propels the vehicle. The agent can harness various tools like a mechanic adjusting the car engine with just the right wrench.
You see, with the typical Language Learning Model (LLM), you feed in a prompt and get a response. Yet, its limitation is seen here: that response is just words. If all you want is a cascade of words, fine, the pure LLM serves you well. However, AI’s agentic nature is meant to do, not just say. When it comes to creating substantial impacts, then the AI tool box comes into play. The AI model requires up-to-date information, maybe just broader data than what the core foundation model offers. The MCP serves as a bridge that connects AI applications to the tool or piece of information they need.
In terms of architecture, the MCP acts as the blueprint of the AI vehicle, detailing how every piece fits together to give a functional machine. The MCP client library forms this host application, similar to the auto mechanic, who has the complete blueprint of how the car engine should look. Tools, prompts, resources, and capabilities are made available and even communicated to the outside world via MCP servers, much like the range of tools available to the mechanic.
Through the MCP, the client and server connect, similar to how different components in a car engine communicate to ensure the vehicle runs smoothly. Interestingly, these connections can either be standard IO, akin to the attendant lines in our imagined engine, or via HTTP and Server Sent Events, sending JSON RPC messages.
Well, how exactly does it all fit and work?
Let’s take an imagined service in charge of making appointments for a tutorial. Your goal is to arrange a breakfast meet-up at a restaurant. Here, we would require some crucial tools and resources: a calendar API integration, data about restaurants in the area, reservation details, and so on.
The process starts with the MCP-equipped application, the ‘host,’ asking the users’ requests. It then throws a prompt to the LLM, which in its basic state might get entangled in a web of uncertainties: “Who or what is a breakfast meet-up? A restaurant?” But with MCP, the limit is lifted. It extends the LLM scope beyond its inherent limitations, letting it use additional resources to provide a better response to the prompt.
So, borrowing from our mechanic analogy, the MCP acts as a universal spanner in our AI garage, fitting and unscrewing various parts as required—pluggable and discoverable. Tools can be turned on and off, added or removed, effortlessly, just as a mechanic would his tools. And much like our imagined engine, where the various components work in a composed harmony, MCP-enabled servers can also function as clients — making them composable and quite sustainable.
So, yes, I hope you can see the larger image now. The Model Context Protocol isn’t just about a standalone desktop application. Rather, it’s a crucial vehicle for building genuinely agentic AI, especially in professional or enterprise settings. It’s a great job for those fascinated with finding ways to make AI a more active participant in the world.
To all the dedicated tech explorers out there, check it out. Roll up your sleeves and get your hands dirty with this fascinating piece of tech. I’d love to hear your experiences and what you manage to build using it.
And remember, these discussions form a small snapshot of the bigger AI and tech panorama. There’s always so much more to explore, to refine, and to rediscover. So keep fanning the fires of curiosity and let it guide you.
