← All articles

How Does an AI Agent Actually Reach Your Apps? A Plain-English Guide to Connectors

October 4, 2026 · 3 min read

You ask an AI assistant to "move my Thursday meeting and email everyone the new time." A few seconds later, it's done. But a chatbot is just a program that produces text. So how did that text turn into a changed calendar and sent emails? The answer is a layer most people never see: connectors. Understanding them is the difference between using agents confidently and handing over keys you didn't know you were giving away.

The model talks; tools act

On its own, an AI model can only read what you give it and write back words. It has no hands. A tool is a small, defined action the model is allowed to request: "search this inbox," "create a calendar event," "read this spreadsheet." The model doesn't click anything. It sends a structured request ("please run create_event with these details"), and a separate piece of software carries it out and reports back.

A connector is simply a bundle of those tools for one app. A calendar connector might include "list events," "create event," and "delete event." When people say an agent is "connected to Gmail" or "connected to Slack," they mean someone has installed a bundle like this and granted it access to your account.

Why there's a standard now

For years, every connection was custom-built. If an AI product wanted to work with ten apps, its engineers wrote ten separate integrations, and none of them worked anywhere else. The Model Context Protocol, or MCP, was introduced by Anthropic in late 2024 to fix that. It's an open standard, a shared plug shape, so an app can build one connector and any compatible assistant can use it.

It caught on quickly. By December 2025, Anthropic reported more than 10,000 active public MCP servers, and support from ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code. That month Anthropic also donated the protocol to the Agentic AI Foundation, a fund under the Linux Foundation, co-founded with Block and OpenAI, so no single company controls it. You'll also hear the older terms "plugin" and "API." They're cousins: an API is how software talks to an app, and a connector is the friendly wrapper that lets an AI use that API.

The part that matters: permissions

Here's the catch. Installing a connector isn't just teaching the agent a skill; it's granting access. Before you click "Connect," check four things:

Read or write? A connector that can only read your calendar is low risk. One that can also create and delete events can change your life in ways you'll notice later. Choose read-only when read-only is enough.

How wide is the scope? "Access to one folder" and "access to everything in your drive" are very different grants. Pick the narrowest option offered.

Who built it? Many connectors are made by third parties, not the app's own company. Treat an unfamiliar one like any software you'd install: check the publisher, and prefer official ones.

Can you revoke it? Know where the "connected apps" or "authorized apps" page lives in the service you're connecting, so you can cut access in under a minute.

There's also a subtler risk. When an agent reads an email or a web page through a connector, text inside that content can try to give the agent instructions. This is called prompt injection, and it's a good reason to keep write permissions tight and sensitive actions behind a human approval.

A five-minute connector audit

Open the connected-apps settings in your AI assistant. For each connector, ask: did I use this in the last month? Does it need to write, or only read? Is the scope broader than the job? Remove what you don't use, downgrade what's too powerful, and keep a short list of what's left. Agents are only as safe as the doors you leave open for them, and now you know where the doors are.

Sources: Anthropic — Donating the Model Context Protocol and establishing the Agentic AI Foundation · The New Stack — Anthropic Donates the MCP Protocol to the Agentic AI Foundation