Before you connect Claude to your inbox, calendar or files, there's one setting worth two minutes of your time. It decides how much control you're actually handing over.
No prompt, no pause. Reserved for actions that can't cause any real harm, like reading an email or checking your calendar.
Claude shows you exactly what it wants to do and waits for a yes. The right setting for anything that sends, changes or shares something.
Off the table entirely, no matter what's asked. Use it for anything you never want an AI deciding on, like permanent deletion.
It's easy to think of a connector as a way for Claude to "look things up" — read your inbox, check your calendar, pull a file from Drive. That's part of it, but it's not the whole picture.
Once connected, Claude doesn't just read your data. Depending on the tool, it can act on it too: drafting an email, editing a calendar event, updating a record, or deleting a file. Some actions are reversible and low stakes. Others are not, which is exactly why the permission setting exists.
The tools will vary depending on what you've connected, but the pattern doesn't change. Reading is low risk. Anything that changes, moves or removes something real is where you want your hand still on the wheel.
You only need to do this once per tool. After that, it stays set until you change it.
You don't need to land on the perfect setup on day one. It's completely reasonable to leave most things on Needs approval to begin with, watch how Claude actually uses each connector for a week or two, and shift the ones you're comfortable with over to Always allow once you trust the pattern.