What your accomplice may do alone
By default your accomplice asks before every action that changes something outside the conversation. You can let it act directly on chosen families of actions.
Turn on a family
- Open Settings → Accomplice → Without asking.
- Each row is a family of actions it would otherwise ask about (shopping items, a records domain, an integration's writes), with the actions it contains.
- Turn on the ones it may do directly, then Save changes.
By default, only shopping lists and their items are on: adding milk to your groceries should not need a card.
Every action taken this way is still recorded in Activity as approved by this setting, so you can always see what happened.
The risk check
A second opinion can send a single action back to you even when its family is allowed. When the Jev risk check is connected (a TypeSafe key in Settings → Extensions → Jev risk check, or ALMA_TYPESAFE_API_KEY), before each action your accomplice may take alone it asks five questions: is it hard to undo, does it involve money, other people, private information, or is it unusual? If any is likely, the action comes back to you as a normal card, with the reason. If the check cannot answer, the action is asked about too.
The check runs at TypeSafe: the action's name and arguments are sent there, finance and health included. Leave it unconnected if you do not want that; your settings then apply as they are.
What never runs alone
- Anything not covered by a family you turned on.
- Sending messages as you in chats (each waits for approval unless you turned on Smart approval for Telegram routine replies; see Telegram).
- Reorganising a vault that is not your accomplice's own.
Earned autonomy
Coming soon — a first week in which your accomplice only shows what it would have done, then proposes taking over one kind of action at a time, from evidence ("you approved all 12 grocery additions — shall I just do these?"). See Goals, initiatives and reflection.