Let’s talk about MCP, the Model Context Protocol standard. I’ve seen countless social media explainers on MCP, but I want to explore the practical side of standards like this, not just the tech. Most of the explanations of MCP frame it as a universal standard for interconnecting AI with… everything, like the USB-C ports on the side of your laptop do for peripherals. These explanations are not wrong, but I think they’re missing the point about why standards are so important. Large Language Models (LLMs) are predictably good at language, and they are getting better at reasoning, but they aren’t great about referencing real-time, factual data. Lots of techniques (like Retrieval Augmented Generation, or RAG) are being explored to add factual context to LLMs, but MCP takes a more strategic view.
Let’s go back to the USB analogy. Before ~2000 it was a lot more technically complex to connect a peripheral to your computer. A keyboard and mouse was pretty straightforward, printers were a bit more complex, but a Zip drive? Joystick? Scanner? You better hope you have enough serial ports, or SCSI ports, or MIDI ports and that you configured the right driver pointing to the right connector labeled obscurely on the back of your computer! The Universal Serial Bus (USB) standard changed all this by creating a plug and play solution that would work universally.
Standardization like this is an important paradigm in the evolution of not just technical architectures, but in the way we use computers. Users shouldn’t have to be experts to connect common components, they should just work. Take business intelligence tools as an example. If you want to create a visualization of your business’ profit and loss, you might use a tool like PowerBI that can create dynamic reports based on a dataset. But what if PowerBI only worked with Microsoft SQL databases? You’d likely need a database engineer and maybe a costly migration project to get your financial data in the right place and format. Thankfully, standards like SQL make this easier, and you can sign into many different data sources from PowerBI.
Let’s get back to LLMs. In my work on Project Charlotte, an AI-driven real estate tool, we gathered a dataset of real estate listings by scraping publicly available data and putting it into a database for use in our application. We then used multimodal LLMs to analyze listings and generate both metadata to enhance our dataset, and human readable reports for realtors. But what if we wanted to go back and add an additional piece of information, say census data for a neighborhood? This required a completely different type of component to request data from the census bureau, an update to our database to store that information, and revised instructions (prompts) for our LLM so we could incorporate that data into our analysis.
MCP offers a simpler solution, just like USB did for peripherals. What if, instead of all this technical setup, we could enter a single URL that just worked? This is the dream of MCP. If we wanted to pull data from the Census Bureau, we could add their address as a data source and then their MCP server would provide clear instructions on how to request data, the formats involved, and instructions for using that data. Our LLM is already great at understanding and translating data formats, so it might already have enough information to take a property address, call the census bureau for additional information, and then incorporate the response into our existing reports. MCP could eliminate the need for custom components and give us plug and play access to data!
It’s not quite that simple today, but the industry attention MCP is garnering reminds me of the early days of USB. Businesses would be wise to adopt standards like this as they roll out AI to maximize future interoperability. I worry a bit about the pace of change, but even early adopters of USB 1.0 can still work with modern devices (my Microsoft IntelliMouse still works great!). Of course, ensuring secure access and proper permissions remains critical, and MCP provides strong provisions for security and privacy as well.
So… you’re MCP curious, what now? For consumers, I’d check out Claude Desktop. This offers a glimpse into how the data interoperability of LLMs+MCP could benefit consumers by offering a new way to accessing and search through your own data using MCP. For businesses, I’d research standards like MCP as you roll out your AI strategy. Even if you’re not ready to implement, it’s helpful to think about modularity and ways to extend your platform in the future. Maybe you decide to ingest some additional dataset to enrich reports for your users, maybe you decide to sell your data to third party integrations. Don’t settle for working in a silo!
As always, if you have questions about MCP, AI, or just want to connect, reach out and I’d love to say hello!
Leave a Reply
You must be logged in to post a comment.