A connector is a link between Claude or ChatGPT and one app you already use
One connector, one app. Gmail is a connector, and so are Notion, Xero, Slack, your calendar and Google Drive, each set up separately.
Once it’s connected, it can look things up in that app and do things in it in the middle of a normal conversation, without you opening the app at all.
One thing that trips people up before they start: Claude calls these connectors and ChatGPT calls them plugins. Same idea, different word, and the rest of this page says connector.
Open one up and you’ll see a list of everything it can do in there
Every connector comes with a list of the individual actions it can take. Worth a look, because it’s the thing that tells you whether the job you have in mind is possible at all.
Six of the Gmail connector’s, on Claude:
Search email threadsFind the thread, the way you would in the search box.Get email messageRead one, in full, including the reply chain.Create a draftWrite one and leave it sitting in Drafts for you.Reply to emailAnswer a thread it has just read.Send email messageActually send it. This is the one to think about.List labelsSee how your inbox is organised, so it can file things where you do.
Claude and ChatGPT don’t carry the same list, and that’s the reason to check rather than assume. Claude’s Gmail connector can’t open an email attachment at all. ChatGPT’s Gmail plugin can. So anything that starts with a PDF landing in your inbox works on one and not the other, and nothing except the list would tell you that.
Claude publishes the full list on the connector’s own page before you connect anything: the Gmail connector is a good one to look at. ChatGPT’s Gmail plugin page tells you who built it and whether it can write, but not which actions it has. To see its available actions, install the plugin first. Then find it on your Plugins page, click the three dots, choose Manage, and scroll past the permissions to Read actions and Write actions.
It reads from your tools, and it writes back into them
Reading is the half people expect. It finds the invoice, pulls up the client’s page, catches up on a thread that’s been running all week.
The writing half is the one people don’t see coming. It drafts the reply into your Drafts folder, adds the task to your tracker with the owner and the date on it, and puts the follow-up meeting in your calendar.
Which means you stop copying and pasting between windows
The old way of working with AI is a lot of copying out of one tool, pasting it into a chat box, then copying the answer back. And it often doesn’t come back in the form or structure the destination tool actually needs. You get something you’re happy with in the chat, then you reshape it by hand, every time.
And the answers get better, because it’s working off the real thing
Good answers don’t come from clever wording, they come from what you feed it. An answer built from what you managed to paste in is an answer about your summary of the situation. Point it at the real inbox, the real project, the real numbers, and you get an answer about the actual thing.
Ask it about a client and it can read the last email, the open tasks and the invoice history before it says anything.
Connecting a tool isn’t handing your data over to be trained on
“I’ve just given AI access to my data” is the reaction almost everybody has, and the worry underneath it is usually a single one: is my stuff going to be used to train it.
Both of them are on the record about it. Claude doesn’t train on anything a connector fetches, on any plan. ChatGPT might, on a personal plan, unless you turn off “Improve the model for everyone”, which is in Settings under Data controls. On Business, Enterprise and Education it doesn’t by default.
One footnote on both. Once something an app has fetched is being discussed in the chat, it’s part of the chat, and your chats are covered by whatever your training setting says. Turn that setting off, or be on a business plan, and none of it is used either way.
You decide what it’s allowed to do without asking you first
It asks before it acts, and how often it asks is a setting you control.
Reading is the quiet half. Nobody sets an approval on a search. Writing is where it matters, and both hold back there by default. Claude sorts a connector’s actions into two groups, the ones that only read and the ones that write or delete, and lets you set each group to Always allow, Needs approval or Blocked. Gmail’s writing group starts on Needs approval. ChatGPT starts on Allow low-risk actions, which lets the harmless ones through and stops for the rest.
Set it on purpose rather than clicking past it. On Claude it’s Customize, then Connectors, then the connector itself, under Tool permissions. On ChatGPT it’s the three dots next to the plugin, then Manage.
Start by reviewing every action until you’re confident it is doing what you expect. Once it is, you can allow the low-risk writing it does reliably. Keep approval on for things that are hard to undo, such as sending, forwarding or replying to an email. A draft you don’t like costs you nothing. An email that has already gone is a phone call.
Connecting a tool takes two minutes, unless someone has to say yes first
- 01Find itCustomize on Claude, Plugins on ChatGPT, both in the left-hand sidebar.
- 02Connect itThe button says Connect on Claude and Install on ChatGPT. You sign in with your normal account for that app.
- 03Check the permissionsLook at the action list while you're there and decide what needs approval.
- 04Ask it something you knowCheck the answer against the real thing before you trust it with something you don't.
If you’re on a paid team plan, there’s a step before all of that, and the two platforms are opposites. On Claude nothing is available until an Owner switches the connector on for the organisation, and even then everyone still signs in individually. On ChatGPT Business most are on by default and the admin’s job is turning off the ones they don’t want. Find out which side you’re on before you decide it’s broken.