Step Up to the All-You-Can-Eat Data Buffet

If you want to know what is going on in the world today, you are not going to read the news reports from last month. But that is effectively the situation we are in when we ask questions of many of the latest and greatest large language models (LLMs). These algorithms are trained on mountains of data, which tweaks their parameters until they are capable of naturally responding to just about any prompt that we can throw at them. But when it comes to knowledge of events that occurred after the training date, these models are going to have nothing more than inaccuracies and hallucinations to offer you.

Training massive LLMs is a complex, lengthy, and exceedingly expensive procedure, so that is not a very practical way to keep these models fresh. A better option is to supply them with a variety of external data sources that they can reference to help them in their reasoning process. This approach comes with some challenges of its own, however. Data is stored in many different locations, using a wide variety of access methods. Configuring and maintaining all of these independent connections can quickly turn into a nightmare for developers that takes far too much of their time that could otherwise be spent improving the product.

Only solutions

One potential solution to this problem has just been introduced by Anthropic. They call it the Model Context Protocol (MCP), and it defines a new standard for connecting artificial intelligence (AI) applications to data sources. Before you get any jitters over the name, I should let you know that I have assurances that this MCP bears no relation to the evil MCP that enslaved poor, unsuspecting programs in TRON.

The MCP — if it catches on with the community — has the potential to replace countless protocols with a single and secure method of accessing data. This open source framework has two primary components — MCP servers, which take on the role of serving up data sources, and MCP clients, which AI applications use to quench their thirst for more data. Together, these components create a complete system in which AI applications can graze on a little data from here, a little data from there, and whatever else they need without requiring any heavy lifting from the development team.

What are you waiting for?

Specifications and SDKs have been released to assist developers in getting started with the MCP. A number of MCP servers have already been stood up for popular utilities like GitHub, Google Drive, and Slack, which makes the initial investment of effort a bit easier to stomach. If you are raring to go, the best way to kick things off is to run through the quickstart guide. Until then, in the words of that other MCP: End of line.