Product AI-agent access
Your WhatsApp workspace, connected to any AI.
Point Claude — or any assistant that speaks MCP — at your inbox and ask it, in plain language, what is happening: busiest chats, who is waiting, what a customer said last week.
- 1Point an assistant at your workspace
- 2It reads — chats, tickets, response health
- 3Scoped to the admin who connected it
- 4Revoke it and the next call fails
Why it exists
The answer is in the inbox. The question is in your head.
“Which accounts have gone quiet this month?” takes a person twenty minutes of scrolling and a machine about a second — but only if the machine can read the inbox. Most tools hand you a dashboard and hope it happens to hold the question you were going to ask.
The reason this is rare is that it is only safe if the connection inherits a person’s permissions rather than the workspace’s. An assistant that can see everything is a data-protection problem wearing a helpful face.
What you get
In the product today.
Twenty tools that only read
Search messages, read a thread, list tickets, check response health, find a contact, look inside shared photos and files. Voice notes are transcribed, so what somebody said out loud is searchable exactly like what they typed.
One that can send, and only just
A single tool replies with one message to one chat. It cannot start a conversation, reach a number that never wrote to you, or run down a list — and it must have read that chat first. Sending is a separate permission a person grants on screen, and it is never on by default.
Scoped to a person, not to the workspace
The connection acts as the admin who made it and sees exactly what they see — private chats, access labels and the history-from date all still apply. Nothing becomes visible because a machine is asking.
You issue it, you revoke it
Only an owner or admin can create a connection, and the consent screen names what is being granted before anything is issued. Revoke it and the very next call fails — the check runs on every request, not once at connection time.
Rate-limited on the way out
The one tool that can send is capped per connection and per workspace, so a loop is refused rather than delivered.
When you would use it
Situations this is for
- “Who has been waiting longest, and what did we promise them?”
- “Show me every chat that mentioned the new price list.”
- A weekly summary nobody has to assemble by hand.
- Asking what was said in a voice note without listening to it.
Questions
AI-agent access — questions
What is MCP?
The Model Context Protocol — the open standard AI assistants use to connect to outside tools. If your assistant speaks it, it can speak to your inbox.
Can the AI send messages to customers?
Only through one deliberate tool, one chat and one message at a time, and only inside a conversation the customer already started. Everything else is read-only. Sending is a separate permission you grant explicitly — a connection that asks for nothing gets read access and nothing more.
Is my data safe?
The connection is scoped to your workspace and to the permissions of the admin who created it, and that is rechecked on every single call rather than once when it was set up. If they lose admin, it stops. If you revoke it, it stops immediately.
Which assistants can I use?
Any that speaks MCP. Claude is the one we build and test against.
Is it on every plan?
Yes — connecting your own assistant is not a Pro-only feature. Creating the connection takes an owner or admin.
Read next
Stop running your best channel off a phone in a drawer.
Connect your first number in five minutes. Keep the number, keep the history, and finally see the whole thing.
No credit card required.