Skip to content

From request to final receipt

After consent, you write an instruction in the chat. The extension obtains context from the active page and sends it with the request to the configured model. Attachments and images can be included. Recent conversation context may be reused: storing history locally does not mean conversation context is never sent to the model.

With Show plan before execution active, a card with the goal and the tabs involved appears before the first call to the model. Steps are added when the model proposes tools; the card is not necessarily a complete forecast of all future actions.

Before the first actions, the plan flow can ask for approval even with Allow all active. A request that only produces text may finish as “Direct answer — no actions”.

The panel uses prices and token counts to show estimated costs. Each round can add cost; see Costs for the limitations.

Additional confirmations depend on the categories enabled in Security, on Allow all, on the domain and on the action’s classification. When a confirmation appears, check the target and the effect before accepting. Refusing blocks the pending action; it does not undo previous actions.

The extension runs the tools on the page and consults the model again when needed. If it finds a sensitive field, it may ask you to fill it manually. You can use Stop or Escape to cancel the turn in progress.

The receipt identifies the result, provider/model, duration, pages, actions, blocks, confirmations, links and available usage. Blocked or cancelled results do not count as a completed task.

The last 50 receipts are stored locally, even if you disable receipt display in the chat. The receipt signature is a local cryptographic digest, not an external certification that the task was fulfilled. You can view the receipts in the panel and delete them in Privacy.