We draw the line, not the provider
“Read my spend” and “change my spend” are different questions, and on several platforms the provider’s own permission screen cannot tell them apart. Google, for one, grants reading and changing Google Ads together, with no read-only option to ask for. So the provider’s permission cannot tell you what your assistant will be able to do, and we would rather say that than pretend.
So we draw the line ourselves. An account you have not opened for changes cannot be changed through HeyMetra, whatever the provider would technically allow.
Where you have opened it, your assistant can only suggest a change. The approval above decides what reaches the account: one change, the one you read, at the moment you say yes.
Questions this page does not cover
If your evaluation needs something specific, such as a particular control, a data residency requirement or a compliance questionnaire, send it to us and we will answer it directly rather than guess at it here. Knowing which requirement matters to you is genuinely useful to us.
Frequently asked
Does HeyMetra need write access to my ad accounts?
Not unless you give it. Every account starts read-only. Some platforms cannot grant reading without writing: Google gives both on a single Ads scope. So we draw the line ourselves: an account you have not opened for changes carries nothing your assistant could use to make one, and an account you have opened still sends every change to a person for approval first.
Can a tool call change something without asking me?
Not in a connected account. A change there is held for your approval, and what you approve is recorded exactly as you saw it; the model does not get to decide again after you approve. A message to the Slack channel, Telegram chat or email address you linked goes straight out, unless you switch on Ask before sending for that connection. An email to an address you did not save always waits for you.
Does HeyMetra store my passwords?
No. Connections that use OAuth are authorised on the provider’s own sign-in screen, so we never see the password. Access tokens and any API key you paste in are stored encrypted, and are never shown back to you or written to logs.
Is my data used to train a model?
Not by us: we do not operate a model. A tool result is read from your account and returned to the assistant you asked in. What that assistant does with it is governed by your agreement with its operator, and that is the agreement to read on this question.
What happens if the model calls the wrong tool?
For a read, nothing lasting: you re-ask and the mistake evaporates. For a write, what you see before anything happens is the current value, the new value and how to reverse it.
Something here not answered? Ask us. Security questions reach a person, not a form.