AI chat on a shared window

AI chat lets anyone in a meeting open a live conversation with an AI model (Google’s Gemini Live) about a window that is being shared. The model sees the shared window and hears the room while a session is live, answers by voice and text, and — only with the sharer’s explicit approval — can act on the window.

It is off by default and is an opt-in feature with real privacy consequences, so read this page before turning it on.

Turning it on

Settings → AI chat → AI chat on shared windows. Flipping the switch on shows the two consequences below and asks you to confirm with Turn on AI chat. Once on, every shared window in your meetings — the ones you share and the ones other people share — gets an AI chat button in its header, and a matching entry in each shared window’s hover-tab and Petal View options menus.

Turning it on also means anyone in your meetings can start AI chat on a window you share. While a session is live, that window’s content and the room’s voice are sent to Google.

By default Petal’s hosted service supplies the AI credentials, with modest limits: a session token is single-use, must connect within 30 seconds, and sessions are capped to a few per person per room per hour. If you’d rather bill your own Google account, paste a key under Gemini API key (optional) (the field appears once AI chat is on) — roughly 2–4¢ per minute of AI chat. Free-tier keys may allow Google to use content to improve their models.

Starting a session

1. On a shared window, click AI chat in its header (viewer side) or choose Start AI chat on this window from the sharer’s hover-tab menu.
2. A floating AI chat panel opens beside the window with the transcript, a text box (Type a message… / Send), a push-to-talk button, and the time left in the session.
3. Everyone viewing the window sees an AI chat live badge on its header for the whole session — “AI chat is live. This window and room voice are sent to Google.” — with their own push-to-talk button, so anyone can speak to the model.

Push-to-talk is deliberate: the model only hears the room while someone is holding the button (“Listening — release to send”), never when humans are just talking amongst themselves. If someone else already has the floor, your talk attempt is refused with a note.

The session runs on the sharer’s machine (that’s where the pixels are), regardless of who started it, and stops when anyone clicks Stop AI chat on this window, the share ends, the time limit is reached, or the sharer disables AI chat.

Letting the AI act on the window

Answering questions never touches the window. If the model wants to click, type, or scroll — for example, “open the second tab” — Petal stops and asks the sharer: The AI wants to act on this window, with Allow once, Allow for this session, and Reject.

• Allow once performs that one action.
• Allow for this session gives the AI standing access to this window until the session ends; the panel shows “The AI has standing access to this window for this session” with a Revoke access button.
• Reject turns window control off for the rest of the session (“Window control is off for the rest of this session”); Allow the AI to ask again re-enables the prompt.

Only the sharer ever sees this prompt, and nothing the AI does bypasses the same Accessibility-based injection that human remote control uses.

If something goes wrong

The panel and header tell you plainly: “AI chat reached its time limit”, “Too many AI chat sessions just now. Try again shortly.”, “The AI chat quota for this key is used up”, “This AI model is unavailable — update Petal”, or “Could not reach the AI chat service.” A message that fails to send stays in the box so you can retry.

Browser participants

The browser client shows the AI chat button and the live badge on shared tiles too, so browser participants can start sessions and talk to the model. The session itself always runs on the desktop app that is sharing the window.