MCP & connectors

Under Settings → Connectors, the model can reach tools you already use — Drive, Gmail, calendars, project trackers. There's no magic in how: a connector translates the tool's data into text the window can read, and your requests into actions the tool understands. The model never "sees" your inbox; it reads a text rendering of it, riding in the window like everything else you've studied here.

The plumbing standard behind this is MCP — Model Context Protocol: an open protocol for describing a tool to a model. An MCP server announces "here is what I can do — search these files, create this task, read this calendar — and here is what each action needs." Any model that speaks the protocol can use any tool that offers it, which is why the same Notion or Linear connection works from the chat app, the command line, and an agent alike: they're all talking to the same server, in the same language.

Two working consequences:

• Connections are capabilities, and capabilities are risk. A connector that can read is a research assistant; one that can write is an actor. Grant reach the way you'd grant it to a new hire — read first, write after trust, and keep an approval gate on anything that leaves the building.
• Try one chore today. Connect one tool and ask for something you'd never bother doing yourself: "find the three oldest unanswered emails from real people and draft replies for my review." The draft-for-review shape is the habit worth keeping.

Say it back in one line: what does a connector actually pass to the model?