The chat experience

Here’s what a conversation in the widget actually looks like for your end users.

Streaming replies

Answers stream in as they’re generated, so there’s no waiting for a full response. When the assistant reasons through something, an optional, collapsed-by-default thinking trace shows its steps — you can turn it off.

Suggestion chips

Action-oriented suggestion chips appear above the composer; tapping one sends it as the next message. As the user types, the widget can also suggest completions drawn from your capabilities.

Action cards

When the assistant wants to do something, it shows it clearly before anything happens:

• An action plan previews multi-step work; nothing runs until the user approves it.
• A confirm card gates a write or destructive action, with a summary of the side effects — and, where required, a typed-confirmation phrase.
• Progress and result cards show what ran and what came back.

Step-up & consent

For sensitive actions, a focus-trapped step-up dialog asks the user to re-confirm their identity (MFA, a passkey, or their password). If you’ve enabled it, a one-time 'act on your behalf' consent is shown the first time and remembered after.

AI disclosures

The widget makes clear the user is talking to AI: a persistent indicator, plus a dismissable banner on the first message. This supports transparency obligations such as the EU AI Act.

Citations, history & attachments

• Citations — answers from your knowledge link to their sources, viewable in a side panel.
• History — past conversations are kept (encrypted on the device) so users can return to them.
• Attachments — users can add images and documents to a message where enabled.

Next steps