August 18, 2026 · By CJ
What Is MCP? How AI Agents Actually Reach Your Apps
MCP, the Model Context Protocol, is a standard socket an app can offer so any AI agent can plug into it. Without one, connecting an agent to an app means somebody wiring that specific agent to that specific app by hand. With one, the app exposes what it can do once, and anything can use it.
The confusion it clears up: having the login isn't the same as being able to reach the app.
A key and a keyhole are two different problems
You can hand somebody a key to your office. If the building has no door, the key is decorative.
That sounds silly because buildings always have doors. Software doesn't. Plenty of apps were built on the assumption that the thing using them would be a person sitting at a screen, clicking. There's a way in for a human and no way in for anything else.
A login is for a person
This is the part that surprises people, and it's worth being precise about.
A login expects hands to type, eyes to read a screen, and a browser to hold the session. An agent has none of those. Giving it the password doesn't help, in the same way giving somebody a key to a building with no door doesn't help.
What an agent needs is a way in built for software: a defined list of what the app can do, what each action needs, and what comes back.
What MCP standardises
Before a standard existed, every connection was bespoke. Your agent and your calendar needed a piece of glue written specifically for that pair. Ten apps meant ten pieces of glue, and each one broke on its own schedule.
MCP makes it one shape. The app offers a socket describing what it can do. The agent reads that and uses it. Nothing in between is custom.
Practically, that changes three things:
- Adding an app stops being a project. If it offers a socket, the agent can use it.
- The app decides what's exposed, which is where your tools boundary actually lives.
- Swapping the agent doesn't break the connections, because the socket belongs to the app.
Where it sits against the other pieces
MCP isn't a fourth thing alongside instructions, tools and triggers. It's the plumbing under tools. Tools are what an agent is allowed to reach. MCP is one of the ways reaching is possible at all.
An app your agent can't reach isn't a tool. It's a screenshot.
The question
Pick the app you most want your agent working inside. Ask whether it offers a socket, or whether somebody would have to build one.
That answer, not the password, is what decides whether this is a five minute job or a project.
What is MCP?
The Model Context Protocol, a standard socket an app can offer so that any AI agent can plug into it and use what it does, rather than needing custom glue written for that specific pairing.
Why can't my AI agent use an app I have a login for?
A login is designed for a person with hands, eyes and a browser session. An agent has none of those. It needs a way in built for software, which is what a socket provides.
Is MCP the same as an API?
Related but not identical. An API is one app's own way in. MCP is a shared shape, so an agent that understands the protocol can use any app offering it without bespoke work.
Does MCP replace instructions, tools and triggers?
No. It is the plumbing underneath tools. Tools are what an agent is allowed to reach, MCP is one of the ways reaching is possible at all.
What does MCP change day to day?
Adding an app stops being a project, the app itself decides what is exposed, and swapping which agent you use does not break the connections.