The assistant: how a conversation works

What the assistant is

Mirra's assistant is a voice conversation with the screen on your wall. You summon it, it listens, it works out what you meant, it uses the household's own data to answer, and it speaks the answer aloud while the face on the glass lip-syncs to it.

It is not always listening. The microphone is opened when you summon it and closed again when the utterance ends — audio arriving at any other moment is discarded before it goes anywhere. That is worth knowing because it explains the one thing people find odd at first: there is no "hey" that works from across the room. You start the turn deliberately.

The conversation lives on the screen you started it on, and it is thrown away after ninety seconds of silence. Follow-ups within that window carry context — "and tomorrow?" works — and after it, the next thing you say starts fresh.

There is no phrase that summons the assistant on any screen today. See the article on naming your assistant for what the wake-word setting actually does, which is less than its name suggests.

Summoning it

1. Tap the microphone button

   It sits centred at the foot of every running screen. On a screen with a keyboard, M does the same thing.

2. Speak

   The face changes to its listening state and your words appear as a caption under it while you talk. The caption firms up as the transcription settles.

3. Wait for the answer

   The face moves to thinking, then to speaking. The reply is spoken aloud and captioned at the same time.

4. Or press T and type instead

   A box appears reading "Ask Mirra…". Press Enter to send. This is the fallback when the microphone is unavailable — the screen says "Microphone unavailable — press T to type instead" — and it is also the quiet option in a room where somebody is asleep.

5. Press Escape to dismiss it

   That stops the microphone, stops any audio playing, and clears the captions. It is also how you put the face back to sleep: once an answer has been spoken the face settles back to its summoned state and waits there. On a screen with faces enrolled, the room emptying dismisses it too.

Tapping the microphone while Mirra is mid-sentence interrupts it and starts listening again, so you do not have to wait out an answer you have already heard enough of.

What happens between hearing you and answering you

Your speech is streamed to transcription as you talk, in British English. When the transcription settles on a final sentence, that sentence — and nothing else — starts the turn.

Before any model is asked anything, Mirra checks that your household is allowed to spend. A refused turn costs nothing and says why. Then the model is given who it is, which screen it is on, the time on that screen's own clock, whether it recognises the person in front of it, and the list of tools it is allowed to use right now.

The model can then go round the loop up to six times: ask for a tool, read what comes back, ask for another, and finally answer. Most turns take one round or none. If it never settles, Mirra says it lost its train of thought and asks you to try again, rather than sitting silent.

The answer is then synthesised into speech, along with the mouth shapes the face uses to lip-sync it. Replies are deliberately kept under fifty words, with no lists and no markdown, because everything here is spoken rather than read.

What it can do for you

These are the actual tools the assistant can reach. Anything not on this list, it answers from its own knowledge or not at all.

• Your diary — upcoming events from every calendar you have connected, a week ahead by default.
• The weather — current conditions and a three-day outlook for the location your weather widget uses.
• What is on the screen — which scene is showing, which widgets are actually painted, why that scene and not another, and the artwork if there is any.
• The artist's own note — plays the recording an artist made about the picture currently on the wall, in their voice, where one exists.
• Tasks — read the open list, add to it, tick something off by a few words of its title.
• Reminders and timers — create a reminder that pops up on the screen at a time, list what is coming, or start a countdown that chimes.
• Scenes — switch the screen to one of its named scenes.
• The home — check the state of lights, doors, locks, temperature and power, and change them: on, off, toggle, open, close, lock, unlock, or set a colour, warmth, brightness or target temperature.
• Music — play a song, album, artist or a mood through Spotify, name a speaker to play it on, and pause, resume, skip or go back.
• Messages for people — leave a message for somebody who lives here, to be passed on the next time a screen sees them.
• Memory — remember a fact you ask it to remember, or forget one.
• The web — look up something that turns on recent facts, if you have switched searching on.

Three of those are conditional. Memory appears only when your household has remembering switched on; music only when Spotify is connected; web search only when you have turned it on — and only where the installation has a search provider configured, which the settings page tells you about if it has not. A tool that is unavailable is left out of the conversation entirely rather than offered and refused — a mirror that spends a whole turn discovering it cannot do something is worse than one that never offered.

What it answers without any tool at all

The tools cover your house. Away from them, Mirra answers as a well-read friend would — history, science, cooking, spelling, sport, how something works, a word in another language. It is told, in as many words, to be genuinely useful rather than deflecting to a search engine, because a household appliance holding a dozen household tools will otherwise lean towards "I can't help with that" for anything outside the front door.

The honest gap is recency. The model does not know what has happened in the last few months, and it is told to say so briefly rather than guess. That is the one thing web search exists to close, and the only thing.

When it has read something from the web

Everything else the assistant reads was written by your household — your calendar, your tasks, your lights. A web result is written by whoever ranked well for a query, and it lands in the context of something that can switch on a light and write to memory.

So a turn that searches stops being able to act. The first search is the only thing that runs in that step; anything the model asked for alongside it is refused with an explanation. From then on the assistant can still read and answer — the diary, the weather, what is on the screen — but every tool that changes something is withdrawn until the conversation ends and somebody speaks to it afresh. The search tool goes with them, so a conversation gets one look at the web and no more.

Because the conversation is what carries the risk, that restriction lasts for the rest of the session rather than the rest of the sentence. If you have just asked Mirra to look something up and then ask it to turn a light off, it will decline — wait for the session to lapse, ninety seconds of quiet, and ask again.

Reading conversations back

Two things about transcripts are worth being clear on. Anyone in your household with full access can read all of them; an account restricted to menus sees none of it. And when Mirra speaks first because a screen recognised somebody, there is no "You" line — nobody asked anything. The brief the portal handed it is stored all the same, and since the transcript marks anything that is not your side as Mirra, it appears above the reply under Mirra's name.

How long transcripts are kept is your choice, under Settings, in "How long we keep things". "Assistant conversations" is set to 90 days unless you change it, and the choices are 7 days, 30 days, 90 days, 6 months, 1 year, or keeping them until you close the account. Once a message is past the window it is deleted for good.

1. Open Assistant in the navigation

   That lands on Voice & appearance, because changing the voice is what people come here to do far more often than reading old conversations — and because a page of past conversations is the most private thing in the product to have on a kitchen wall by default.

2. Choose Conversations

   The button sits at the top right of that page. It lists the last thirty sessions, newest first, each labelled with the screen it happened on, when it started, and how many messages it holds.

3. Click a session to open it

   Every line is shown in full: your side marked You, Mirra's side marked Mirra.

What a conversation costs

The assistant is metered in credits. One credit is one token — a word or so of what goes into the model plus what comes out of it — and your plan carries a monthly allowance of them that resets on the 1st.

Speech is deliberately outside that. Transcribing what you said and synthesising the reply both cost us money and are both recorded, but neither is charged against your credits, so the number you are watching only moves for thinking.

• A conversation is roughly 700 credits. That is the working figure the billing page uses to tell you how many more you have this month.
• A web search is charged flat at 3,000 credits, whatever it finds. It is a paid call to a third party, which is why it costs more than a sentence of thinking.
• Longer conversations cost more per turn. Everything said so far goes back to the model each time, so the tenth exchange in a session is dearer than the first. Sessions expire after ninety seconds of quiet, which keeps this in check on its own.
• A refused turn costs nothing. The allowance is checked before any model is reached.

When the month's allowance is gone, the assistant stops answering and says so on the screen. Nothing else in Mirra stops — scenes, widgets, schedules and screens carry on exactly as before. Watch it on Billing, under "Assistant credits this month".

Limits worth knowing

• It transcribes British English. The language is fixed in the code; a strong accent in another language will come back as approximate nonsense, and there is no setting for it.
• It hears one screen at a time. A conversation belongs to the screen it started on. Asking the hall mirror a follow-up to something you asked the kitchen one will not work.
• It does not know who is speaking unless a face has been enrolled. With nobody recognised it is told, explicitly, not to guess or use a name.
• Replies are short by design. Under fifty words, no lists, no markdown, no emoji. A spoken paragraph is a paragraph nobody listens to.
• If the cloud provider turns it away, it degrades rather than dies. When the model is not available to the account — access not granted, a key rejected — a simple built-in matcher takes over and says its cloud voice is offline. It can still handle the calendar, tasks, reminders, timers, the weather, the screen and the lights, and it sounds noticeably thinner. It retries the cloud every ten minutes.
• Menu-only accounts see none of this. The assistant pages, the transcripts and the settings all require full access.