Chat and voice

Text and voice conversations with the assistant, including chat history, mid-turn steering, and parallel requests running as background tasks.

The assistant is available everywhere in the app through a chat box: on Mission Control alongside the briefing, on the Chat page for focused conversation, on the Mail page as a composer, and on the Calendar page for scheduling. All of these surfaces share the same conversation thread, so context from a question asked on Mission Control is still available when continuing on the Chat page.

Text chat

Starting a conversation

Open the Chat tab or use the chat box on any page. Type a message and press Send. The assistant streams its response as it works, and shows progress labels when it's doing something in the background (searching mail, checking the calendar, creating a task).

Suggestion buttons offer quick starting points on pages that support them. On the Chat page, these include "What's new in my inbox?", "Am I free Thursday afternoon?", and "Add a task to renew the insurance by Friday."

What the assistant can do

The assistant has access to connected email, calendar, tasks, notes, and web search. Some examples:

  • Email. "What did Kate say in her last email?" "Draft a reply to Mark's proposal." "Search my inbox for receipts from last month."
  • Calendar. "What does my week look like?" "Book lunch with Sarah on Thursday at 12:30." "Find me an hour for exercise tomorrow." "Do I have any clashes coming up?"
  • Tasks. "Add a task to call the accountant by Friday." "What's on my task list?" "Mark the insurance renewal as done."
  • Notes and memory. "Remember that the wifi password is bluefish42." "What's the wifi password?"
  • Web search. "What's the weather in Melbourne tomorrow?" "Look up the latest exchange rate for AUD to USD."
  • Email drafting. When the assistant drafts an email, it goes to the Approvals page for review before anything is sent.
  • Approving a send, hands-free. "What's waiting to go out?" "Read me the draft to Paul." "Yes, send it." The assistant runs the approval itself, so a send can be authorised by voice with no screen to tap. It only ever acts on a direct instruction from you.

New chat and history

New chat starts a fresh conversation. The previous one is saved and available in History.

History opens the full archive of past conversations. Each conversation can be opened to read the transcript, continued where it left off, downloaded as markdown, or deleted.

Export on the Chat page downloads the current conversation as a markdown file.

Mid-turn steering

If the assistant is working on something and the direction needs to change, send another message while it's still responding. This is a barge-in: the assistant stops what it was doing and pivots to the new instruction. Useful when a response is heading in the wrong direction or the question has changed.

Parallel requests

On the Pro and Internal plans, the assistant can work on several things at once. Ask for a few separate jobs in one message and it hands each one to a background task, answers straight away, and reports back once they have all finished.

"Draft an email to Kate about Thursday, scan my inbox for the insurance renewal, deconflict my calendar for next week, and process this business card."

Not every message is split. The assistant decides whether a message holds work that can run in the background; a simple question or a single job is answered directly, as before. On other plans every message is answered directly.

The acknowledgement

When the assistant dispatches work, it replies in one line that names every task and its order, for example: "On it — 4 background tasks: (1) Draft the email to Kate; (2) Scan the inbox for the insurance renewal; (3) Deconflict next week's calendar; (4) Email the attendees (after 3). I'll report back when they're done."

Request cards

The tasks appear as a Background tasks card beneath your message, with one row per task. The card updates live, without a page reload. Each row shows the task and its state:

  • Queued — waiting to start, either for a task it depends on or for a free slot.
  • Running — working now.
  • Retrying — an attempt failed and it is trying again. The row shows the attempt number.
  • Blocked — it won't start, because a task it depends on failed.
  • Done, Failed, or Cancelled — finished.

Tap a finished task to read its result, or the reason it failed. Results land on the card as each task finishes, so the first answers are readable before the rest are done.

Dependency chaining

When one task needs another's result, the assistant chains them: "Email the attendees" waits for "Deconflict the calendar", and the row reads "Waiting on Deconflict the calendar" until it does. The later task starts as soon as the earlier one is done, and it is handed the earlier task's result. Tasks that don't depend on each other run side by side.

A task that fails is retried automatically, with a growing pause between attempts. If it still fails, it is marked Failed and every task that depends on it is marked Blocked.

Cancel, redirect, and retry

Each row has its own controls, and each one touches that task only:

  • Cancel stops a queued, running, retrying, or blocked task. Any task that depends on it is cancelled too, and the confirmation names those tasks before anything happens. Other tasks keep running.
  • Redirect sends a running task an extra instruction ("also copy in Paul"). The task picks it up at its next step.
  • Retry re-queues a failed task. It resumes from where it stopped, so steps it already completed (a draft it already queued, say) are not repeated, and the tasks that were blocked behind it are queued again. On a blocked task, Retry retries the failed task it is waiting on.

The same controls are available by asking. "What's running?" lists the tasks in flight and the most recent finished ones. "Cancel the business card one", "For the email to Kate, also mention the venue", and "Retry the calendar one" each act on the one task named. If a description could mean more than one task, the assistant asks which before doing anything.

The summary

Once every task in a request has finished, failed, or been cancelled, the assistant posts one reply beneath the card that pulls the results together. If something didn't get done, the reply says which tasks and why, and offers to retry them.

That summary is the only notification a request sends. Individual tasks finishing do not notify. If you're away from the chat when the summary lands, one push notification arrives with it.

Several chats at once

While background tasks run, the chat stays usable: send another message in it, start a new chat, or reopen an older one and carry on. Both chat lists (the Chat page sidebar and History) mark a chat that is working with a Running badge, or with a count such as 3 running when background tasks are in flight.

There are limits, set by an administrator:

  • Chats working at once. By default, up to 5 chats can be working at the same time. A message in a sixth chat waits, its chat shows a Waiting badge, and the composer reads "Waiting for a free slot — starts automatically." It starts by itself as soon as another chat finishes. A chat that is already working takes no extra slot, so a new message in it starts straight away.
  • Tasks running per chat. By default, up to 10 background tasks run at once in one chat. Any more wait as Queued and start in order as earlier tasks finish.

A message that has waited for a free slot for 30 minutes stops waiting and says so. Send it again once another chat has finished.

Voice conversations

Voice lets you talk to the assistant hands-free. Speech is transcribed in real time, the assistant responds in text, and the reply is spoken back aloud.

Starting a voice session

On the Chat page, press the microphone button next to the text input. The first time, the browser asks for microphone permission. Once granted, a voice session opens and the assistant is listening.

Speak naturally. The assistant waits for a pause of about three seconds before it treats the speech as a complete turn and begins working on a response. While it's working, it speaks a short acknowledgment so there's no dead air. If work takes longer, it gives periodic spoken updates until the response is ready.

Barge-in

Start speaking while the assistant is still talking and it stops immediately, discards the interrupted response, and focuses on the new input. This is the same mid-turn steering as in text chat, but triggered by voice.

Background tasks by voice

Parallel requests work the same way on voice. Ask for several things in one breath and the assistant speaks its one-line acknowledgement, then is free for the next request. While tasks are in flight, the voice dock shows how many are running, for example "Listening · 3 running".

  • "What's running?" The assistant tells you what's in flight, what's waiting on what, and what has finished.
  • "Cancel the business card one." Cancels that task, and anything waiting on it.
  • "For the attendee email, also mention the parking." Redirects that one running task.
  • "Retry the calendar one." Retries a task that failed.

When the whole request is finished, the assistant speaks the summary, provided the voice session on that chat is still open. It waits for a quiet moment, so it never talks over you or over another reply. If the session has closed by then, the summary is in the chat and the usual single push notification applies.

Echo filtering

The assistant's own spoken replies come through the device speaker. Echo filtering prevents it from hearing its own voice and responding to itself. This works automatically; no headphones are required, though they can help in noisy environments.

Voice settings

The assistant's speaking voice can be changed in Settings → Preferences. A preview button plays a sample of the selected voice before committing. The assistant's name (displayed in the UI and used in voice greetings) can also be set there.

Voice is available on plans that include it. If voice isn't part of the current plan, upgrading to a plan that includes it enables the feature immediately.

Tips

  • The assistant remembers context within a conversation. "What about next week?" works after asking about this week's calendar, without restating the question.
  • Use "New chat" to reset context when starting a different topic. Previous conversations are always available in History.
  • The assistant shows inline maps and images in responses when they're relevant to the conversation.
  • On mobile, the app works as an installable PWA. Add it to the home screen for a full-screen experience without the browser address bar.