Letting AI Interface with your App with MCP

In plain English

Kent starts with a film scene where Tony Stark talks to his AI assistant and it pulls data from many sources. He asks why we do not all have that yet. His answer is integrations: someone has to wire the assistant to every service, and no company will do it for every small service.

The Model Context Protocol (MCP) is a shared standard for that wiring. A service builds one MCP connection, and any assistant that supports the standard can use it. He demonstrates this with his own journaling and weather services. The assistant looks up his location, checks the weather and writes a journal entry, after he approves each step.

Three ideas worth keeping

• The blocker for a useful AI assistant is the effort of connecting it to each service one by one.
• With a standard like MCP, the service owner decides what the assistant can do, instead of waiting for an assistant maker to build it.
• Login and permissions still apply: his demo used a secure sign-in, and he approved each tool the assistant used.

What it means for your business

Customers may increasingly ask AI assistants to do things for them, such as find information or place a booking, instead of visiting your website. A standard connection lets those assistants work with your systems on terms you set. Kent says the assistant apps were not fully ready at the time of recording, so this is early.

What it means for software teams

Building to a shared standard means one connection serves many assistants, rather than a custom integration for each. Kent also notes the server can reply in plain language and the assistant can adapt it, for example translating into the user’s language. Approval prompts for every action are, in his words, a sign that trust is not built yet.

What we found, and what we did

We have not published anything specific on MCP. Our guide How to Connect Your Business Systems So Data Flows and our note The Integration Layer Your Business Is Missing argue for one well-designed hub, rather than a web of one-off connections.

Kent’s point is a similar idea applied to AI assistants. We would watch for who controls permissions and what data the assistant can reach, which our guide to building AI agents also treats as a design decision.

Worth watching

• 04:54: why we do not already have a Jarvis-like assistant: integrations.
• 06:57: his three phases, from chatbot to tool use to a shared standard.
• 13:35: the live demo of location, weather and journal services working together.