Skip to main content
The Agent does nothing that goes out unless you explicitly say yes. It presents the action, you decide.

What always needs a confirmation

  • publishing or scheduling a post via Buffer
  • sending a message in Discord
  • sending a contract via Documenso
  • creating, changing or deleting chat commands
  • putting a link or partner on your SITE
  • adding a public FAQ for the Twitch chat
  • changing the target of a short link
Creating short links, writing drafts, saving QR codes and remembering things it does without asking. Nobody outside sees any of that yet.

What it looks like

A card Wartet auf deine Bestätigung (waiting for your confirmation) appears in the chat. Below it is exactly what should happen, with all details, e.g. channel and text of the post.
1

Check

Read through the details. If something is wrong, tell it instead of confirming.
2

Decide

Ausführen (run) or Abbrechen (cancel). Afterwards the card says Ausgeführt (done), Abgebrochen (cancelled) or Fehlgeschlagen (failed).
The rules:
  • Only the person who triggered the action can confirm it.
  • The card is valid for 15 minutes. After that it says the action has expired, and you simply ask again.
  • When you confirm, oxom checks once more whether you are allowed to do this right now.
On WhatsApp you reply with “yes” or “no” instead, see Via WhatsApp.

After outside content it asks more often

If the Agent has read something from outside in a reply, e.g. a website, a file, a Notion document or your calendar, it asks before every action for the rest of that reply, even before a short link or a draft. Outside texts can contain hidden instructions. This way nobody can make the Agent do something you don’t want through a prepared file. In these cases the card shows the technical name of the action, e.g. shortenUrl, with all details after it.

What gets logged

Under Einstellungen → Logs (Settings → Logs) admins and owners see, among other things:
  • which actions the Agent ran after confirmation
  • which actions were blocked
  • errors in individual steps
  • messages that looked like an attack on the Agent
  • reported missing capabilities
Last modified on October 30, 2026