What is an MCP?
Model Context Protocol or "MCP" is an "open-source standard" for interfaces between applications and AI (predominantly Large Language Models aka LLMs). To break these statements down further, "open-source standard" has two key parts.
The first, "open-source" in this context just means that the template for creation of these interfaces is 'open' and you can go and look at it for free, even use it to make your own protocol! The second, "standard" means that the design of this template, this 'protocol', is a standard that exists across LLMs (though not all). If you're using one model and want to use an MCP developed somewhere else, you can be confident that the created instance will be in the same format.
OK, cool, what does "interfaces between applications" mean? Well, the protocol is like predefined grammar for a blank language. An instance of this language will differ depending on the application, it will have different vocabulary (functions in the MCP) but the grammar will always be the same. In addition, MCPs contain "Resources" which is data (including explicit documentation) explaining what each piece of vocabulary added is and does. Resources can also include files as well. Finally, MCPs include "prompts" which are like playbooks instructing the LLM how to make certain sentences. For example in the language analogy a prompt might specify that in this integration all sentences must end with a certain word as part of best practices, or that the LLM must ask permission of the user before executing any commands in all cases. Overall, this grammar and these patterns have been designed to be maximally understood by large language models. So regardless of the vocab, resources, or playbook you put in your own MCP the LLM that uses it will be able to work out where to look to see the documentation you've included explaining the words you've added and also any common rules it needs to remember when operating.
Hopefully that makes sense. Putting it all together, the model context protocol is a template for MCP instances which can be created to serve as interfaces between an application and a large language model.
What does an MCP actually do?
Interfaces created using the model context protocol are confusingly referred to as "MCPs" despite being created from the MCP, but this is now convention and is something we all just have to get used to.
At their core, MCP integrations expose tools (functions), resources (data), and sometimes prompts (structured workflows) that the model can use to do stuff. In practice this "stuff" can be whatever you want, but in the same way it's perfectly possible to define a word to mean whatever you want it to mean, it's not very useful to do so.
What are MCP "servers" and "clients"?
MCP follows a client–server architecture. The AI application (host) contains an MCP client, which connects to one or more MCP servers. The servers provide tools, data, or capabilities, while the client handles communication between the model and those servers.
Conclusion
MCPs aren't particularly complicated, conceptually, but unfortunately due to some difficult communication trends talking about them with people who aren't already immersed in the AI community has become near impossible. This is unfortunate! Education is vital for us all to make informed decisions regarding the AI and AI tools that have come to dominate our business and even society.
