Approvals in the chat
Approvals in the SovrGPT chat: before a connected service runs an action, a card shows every detail. It runs exactly once, only via the card, for up to one hour. With approval modes and a retention period.
Through connected services the model can trigger actions in the chat, for example filing a document, submitting an application or sending a message. Actions that need an approval only run once you approve them on a card in the chat.
Text in the chat is not an approval. If you write "yes, send it", nothing happens. Actions only run via the button on the card.
When you are asked
- Connections that require confirmation: for Marketplace
templates that can write, SovrGPT asks before every write action, for example before
fit_antrag_sendenwith GovBridge. - Mailbox and calendar that you connect with your own sign-in: every action that does not declare itself as read-only asks. While a mailbox is connected in the chat, actions of other connections ask too, except purely reading ones without outside access.
- Deleting actions always ask.
Custom MCP servers that you add under "Eigenen MCP-Server hinzufügen" (add your own MCP server) do not ask (see Marketplace).
The card
When the model calls such an action, its answer ends with the card "Freigabe erforderlich" (approval required). It shows:
| Item | Content |
|---|---|
| Action | the name of the action, in the card's title |
| Connection | the connection it runs through |
| Details | every value the action would run with: recipient, subject and message at the top, everything else in full below |
| Validity | how much longer it can be approved, at most one hour |
| Status | waiting, running, done (with a preview of the result), declined, expired or failed |
The card shows the details SovrGPT stored when the action was called. It runs with exactly these details, not with a later version from the chat. If the details are too long to show in full, the card cannot be approved.
How it works
- Check: look at every detail. If the request came from a document or a web page the model read, check the recipient and the target with particular care.
- "Freigeben & ausführen" (approve & run) or "Ablehnen" (decline).
- Approving runs the action exactly once. A second click, a second tab or sending again does not run it a second time.
- Declining never runs anything.
- The chat continues: if the card is in the latest answer, the chat continues on its own afterwards. The model learns whether the action ran, was declined or expired, and, if it ran, the result.
Every action needs its own approval. An approval does not cover later calls of the same tool or other details. If the model changes the details afterwards, a new card appears.
| Case | What happens |
|---|---|
| An hour passes without a decision | The card expires. Ask the model to request the action again. |
| The connection has been switched off in the meantime, or the action is gone | Nothing runs. The card says so; once it is switched back on you can try again within the hour. |
| The sign-in with the service has expired | Nothing runs. The card links to "Verbindung erneuern" (renew connection). |
| The run is cut off | The card says it is unclear whether the action went through. Check this in the target system before you request it again. SovrGPT does not repeat it by itself. |
Only the person whose chat it is can approve or decline.
Approval modes
You set when the chat asks under Einstellungen → Freigaben (settings → approvals,
/settings/freigaben):
| Mode | Effect |
|---|---|
| Jede Änderung bestätigen (confirm every change, default) | Every action that needs an approval waits for yours. |
| Nur Riskantes bestätigen (confirm risky actions only) | Reading actions run directly. Anything that writes, deletes or says nothing about itself waits for your approval. |
| Vollautomatisch (fully automatic) | Actions run without asking. |
- The cases from When you are asked ask in every mode, including "Vollautomatisch": mailbox actions, actions of connections that require confirmation and deleting actions. At the moment every action that needs an approval falls into these cases; the mode takes effect as soon as more tools are added.
- Organisation default: owners and admins set, on the same page, what applies to everyone who has not chosen a mode of their own. With "Festsetzen" (fix it) nobody may choose a looser mode; stricter personal settings remain possible.
- Every change to your own mode and to the default is recorded in the organisation's audit log.
What is stored
- The record: for every card SovrGPT stores the details, the decision, the time and, after a run, the result. Only you see your records, not even owners and admins.
- Retention: records are deleted in full 90 days after the request. If you leave the organisation, your records there are deleted immediately. The chat itself is kept like any other chat.
- Audit log: every approval and every decline is recorded in the organisation's audit log, with the connection and the action, never with the details or the result. For mailbox connections only the kind of action is recorded, not its name.
- Call log: the run appears like any tool call under "Letzte MCP-Aufrufe" (recent MCP calls) in the Marketplace.