In talking to people who do not explicitly work in tech, I have noticed that many of them do not know about the Model Context Protocol (MCP). An MCP is an open standard which gives Large Language Models (LLMs) a universal way to connect to databases, tools, and workflows. Connect one to Gmail and the model can read your emails, write new ones, and work inside the tools you already use.
The standard part is what makes this matter. Before MCP, connecting a model to a tool meant building that connection by hand, one at a time. Ten models and ten tools means a hundred separate integrations, and every one of them breaks on its own schedule. MCP turns that hundred into twenty. You write one server for your tool, and every client which speaks MCP can use it.
LLMs are also limited in a way most people do not think about. They cannot store infinite memory, so everything a model can see at one time has to fit inside a fixed space called the context window. When that window fills up, older material falls out to make room for whatever is new. You could've told Claude your name at the beginning of a session, but by the end of it, the model may have already forgotten.
Forgetting is annoying, but it is often easy to catch. The more serious problem has to do with the referent, the actual thing an answer is supposed to pull from. A hallucination is the model still answering after that referent is gone, or when there never was one to begin with. That is a different failure from forgetting.
AI hallucinations are the result of next-token prediction producing fluent text, writing that reads smoothly and sounds confident, whether or not a referent exists. Basically, the LLM tries to answer your question or problem with or without knowing what it should be basing its output off of. Recently I asked Claude to research restaurants which would be optimal for a large 10-15 person work dinner. Claude responded with a list of restaurants, talking about them as if they were the best, most popular restaurants in San Francisco. Something didn't feel right; I was not familiar with any of them even though I have been a San Francisco resident my whole life. I asked Claude where the information came from, and it revealed to me that the hyped up descriptions had no source behind them at all. It gave me an answer without an existing referent.
Catching a hallucination is easy when it’s something you are deeply familiar, but it gets more difficult when the LLM uses your own pattern recognition against you. To understand how LLMs trick you, I want to compare AI hallucinations to logical forms, because the form of correctness does not truly represent correctness.
Validity is the logical form in which, if the premises are true, then the conclusion must be true. Arguments can follow this structure and appear to be true, but in reality one of the premises is false, making the entire argument unsound. Soundness builds on validity, except the premises must be true in the real world for it to be sound. It can't just have the form of truth. To understand this distinction I propose two examples:
Validity can be explained by a claim such as:
1. You work hard
2. If you work hard, you will be rewarded with more money
C. You will make more money
While this argument feels intuitively true and is a common claim made by those who work hard, working hard unfortunately does not necessarily mean you will make more money. This argument is then valid but not sound.
Soundness can be explained by a different claim:
1. My boss sent an email on August 2nd
2. If someone sends an email, it lands in their sent folder
C. My boss's email is in the sent folder
This one is sound, and the reason why matters more than the structure does. Both premises are actually true, and I know they are true because I can go and look. I can open the sent folder and check. That is the difference between an argument which merely holds together and one which holds up.
This is what a hallucination actually is. It is a valid but unsound case. The answer has the shape of a correct one, the reasoning runs cleanly from premise to conclusion, and somewhere underneath it a premise is false or was never there at all. Which means the fix is not to make the model reason better. The reasoning was already fine. The fix is to give it premises it can check.
Remember how I said that AI hallucinations are the result of the LLM outputting an answer even if the referent doesn't exist? Well, what if the referent was always available?
By using MCP connectors to services like Gmail, HubSpot, or Obsidian, you have a referent which lives and is preserved outside of the context window, whereas a referent that exists only inside the context window is gone the moment it falls out of that window. MCPs are not perfect, and a lot of them are relegated to specific domains. Gmail only gives email context and HubSpot only provides customer record context. The way around that is to build a knowledge base of your own and connect that too, so the model has one place to check which covers your whole operation instead of one slice of it.
My team and I tested this at our Q3 company on-site. We used multiple MCP connectors to record the discussions and turn them into a deck summarizing what we should take away from the day. The deck came back accurate, which is what we expected. What we did not expect was that it corrected us. Several things we had said out loud during the day did not line up with what was already in our internal knowledge base, and it caught them. It also filled in details nobody could remember. The record beat the memory of the people who were actually there.
Now that you know what an MCP is and why it matters, go use one. Connect Claude to Gmail, Google Calendar, HubSpot, or a Fireflies meeting recorder, and give it something to check itself against. It will not make the model right about everything, but it will stop it from inventing answers about your own work, which is most of what you were going to ask it about anyway. And because it is a standard, every connection you add after the first one is easier than the last.