Permissions and safety
Exactly what Claude can and can't do in your ActuallyCare CRM, what you approved on the consent screen, and how to disconnect at any time.
Handing an AI the keys to your client database is a fair thing to be nervous about. This page spells out, in plain English, what Claude is allowed to do, what it can never do, and how to shut the connection off in seconds.
The consent screen — what you actually approved#
When you connect Claude to ActuallyCare, a page titled "Grant Claude access to ActuallyCare" appears before anything is shared. (If you weren't logged into ActuallyCare, you log in first, the normal way.) It states the access in four bullets:
- Read your CRM data — escrows, listings, clients, leads, and calendar
- Create and update records when you ask
- Draft emails and texts for your approval — nothing sends without your okay
- Archive or delete records, with confirmation
That's the real grant, stated plainly: approving connects Claude to your CRM as you — it can read your records and create or update the ones you ask for, like opening an escrow or rescheduling an appointment. The bullets are one half of the safety story; the other half is how the tools themselves are built:
- Messages start as drafts. Emails, texts, and call scripts go nowhere until you've read them.
- The one send needs an explicit yes. The only thing Claude can actually send — a review-request email — requires you to clearly say "yes, send it" in the chat first, after you've seen the draft.
- Deleting requires confirmation. Destructive tools are flagged so Claude checks with you before using them.
- Archiving beats deleting. When something needs to go away, Claude prefers archiving, which is reversible — you can restore an archived record any time.
Nothing is shared until you approve this screen, and you can take the access back at any time (see how to disconnect below).
What Claude can do — and what it can't#
| Claude CAN | Claude CANNOT |
|---|---|
| Read your escrows, listings, clients, leads, and calendar | Send anything without your okay — everything starts as a draft |
| Create and update records you ask for — leads, clients, escrows, appointments | Delete things quietly — deleting requires a clear confirmation from you, and most "removals" are reversible archives |
| Draft emails, texts, and call scripts for you to review | Touch your billing, subscription, or payment details |
| Run reports — commissions, production, dashboard stats | See other agents' private data — it works inside your account, with your permissions |
One exception to "draft only": Claude can send a review-request email — but only after it shows you the draft and you clearly say "yes, send it" in the chat.
The pattern to remember: Claude looks things up and writes things down, but the send button and the delete button stay with you.
Your data stays in your CRM#
Connecting Claude doesn't copy your database anywhere. Your clients, escrows, and leads stay right where they've always been — in your ActuallyCare account. When you ask a question, Claude reads the answer through a secured connection, the same kind of protected channel your bank's app uses, and only the pieces relevant to your question pass through.
How to disconnect#
The off switch lives in Claude's own settings, and it works instantly.
-
Open claude.ai and go to Settings → Connectors.
What you should see: a list of your connectors, with ActuallyCare among them.
-
Remove the ActuallyCare connector.
What you should see: ActuallyCare disappears from the list.
That's it — the moment it's removed, Claude's access is cut off. There's no separate page inside ActuallyCare you need to visit; removing the connector in Claude is the whole job. Your CRM data is untouched, and you can reconnect any time by running the quickstart again.
You won't be nagged to re-approve#
The connection keeps itself fresh, securely. Behind the scenes, Claude uses a temporary digital pass that expires every 24 hours and renews itself automatically. You approve the consent screen once, and the connection just keeps working — no weekly "please log in again" interruptions. If it ever does ask you to reconnect, that's unusual; the troubleshooting page covers it.