← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

The Claude Setting
Everyone Should Check

You are probably giving Claude more permission than you realize. Before you connect it to Gmail, Google Drive or Slack, do this one thing first. This is part 4 of my AI safety series.

Most people connect Claude to their tools and never check what they just handed over. I did the exact same thing at first.

A connector is a door into real data, like your inbox, your files and your team chat. And it can act, not only read. In Gmail, Claude can now send, reply to and forward email. In Google Drive, it can share, move and trash files. (Anthropic’s Google Workspace connector docs)

The good news is that Claude doesn’t make this all or nothing. Every connector has Tool permissions, so you decide what each action is allowed to do. It takes about a minute per tool.

The Three Levels

01

Always allow

Claude never asks. It just does the action. I use this for things that only look, like searching my inbox or reading a file. Nothing there can cause damage.

02

Needs approval

Claude checks with you every time before it acts, and you say yes or no. Use it for anything that changes, sends or shares. It costs one click, and that click is worth it. Anthropic’s Gmail and Drive actions start here by default.

03

Blocked

Claude can’t do that action at all. Use it for anything you never want an AI doing on its own, like trashing files or sharing them outside your company.

How I Set Mine

My pattern: Claude does the safe half of a job freely, and I keep my hand on the part that can’t be undone.

  1. Read and search Google Drive: Always allow. Reading doesn’t change anything.
  2. Trash, move or share Drive files: Blocked. An AI has no reason to clear out or share my work on its own.
  3. Draft emails: Always allow. A draft just sits there until I look at it.
  4. Send, reply or forward: Needs approval. I read every message before it leaves my inbox, no exceptions.

My Rule Of Thumb

Reads can lean toward Always allow. Anything that deletes, sends, shares or pays belongs on Needs approval or Blocked, because those are the actions you can’t take back.

Set It Up In A Minute
  1. Open your connectors. In Claude, go to Customize (renamed Plugins on newer accounts) and open Connectors.
  2. Click one connector, like Gmail or Google Drive, to see its Tool permissions.
  3. Set each action to Always allow, Needs approval or Blocked. You can set a whole category of tools at once or go tool by tool.

Do this the moment you connect a new tool, before you let Claude start working in it.

On a Team or Enterprise plan, your company’s owner may have already set these for everyone, and you can’t loosen them. Blocking an action in Claude also never gives it more access than your account in that app already has. (Anthropic’s connector docs)

Audit Your Setup

Not sure which actions your connectors even have? Paste this into a new Claude chat with your connectors turned on. It sorts every action into a level for you, then you go set them.

Connector Permission Audit
You are my AI safety reviewer. I'm about to let Claude work inside my connected tools and I want every permission set on purpose, not by default. My connectors: [list them, e.g. Gmail, Google Drive, Google Calendar, Slack, Notion] What I use Claude for with them: [e.g. drafting replies, finding files, summarizing threads] Things I never want an AI doing on its own: [e.g. sending to clients, deleting, sharing files outside my company] Do this: 1. For each connector, list the actions you can take with it in this account (read, search, draft, send, create, edit, move, share, trash or delete). Only list actions you actually have tools for. If you are not sure a tool exists, say so instead of guessing. 2. Sort every action into one of three levels: Always allow (only looks, changes nothing), Needs approval (changes, sends or shares something), Blocked (can't be undone or I said never). 3. Flag anything that sends data outside my account or can't be reversed, and explain the risk in one plain sentence. Output: one table per connector with columns Action | Level | Why. Then a short checklist of the exact settings I should change, in order. Rules: when in doubt, pick the stricter level. Don't take any action in my tools while doing this review.

The One Rule

The safest AI setup is the one you control. Set your permissions before you connect Claude to anything that holds real data.