Say “Minda” and talk. Nova asks back, works across your apps and answers out loud, and it asks before it pays.
Example: names, numbers and prices are made up
How it works
Four beats, every request.
The shape of light is Nova, one of the agent's eleven looks. It takes the form of whatever it is doing, so you can tell at a glance.
01
It hears you
Say “Minda”, then what you need, the way you would say it to a person. Each stretch of speech, a clip of at most 12 seconds, goes for transcription: to OpenAI with your own key, or to Minda's hosted models on your credits.
Voice
Minda, book a table for two on Saturday, somewhere quiet.
02
It thinks
Short commands, like switching tabs or opening notes, run at once. A job of several steps goes to the SI agent, which plans it and shows the plan as a checklist in the chat.
Working on it2 done
Found 3 quiet places with a table at 19:30
Picked the one with the best reviews
Booking a table for 2
03
It does the work
In your signed-in accounts, in tabs of its own: mail, sheets, chats, calendars, shops. Up to three jobs run at once, each reported back when it ends.
It answers aloud in the voice and language you chose. Before it pays, places a call or shares a file, it shows the step and waits for your Allow. Talk over it and it stops at once.
Booked: Saturday at 19:30, a quiet corner table for two. You pay at the restaurant.
Listening
A conversation, not commands
Talk to it the way you talk to a person.
Nova asks when it needs to know, stops the moment you talk over it, and checks with you before anything costs money.
Talk over it
Start talking and it stops at once, then takes what you said as the next request. No button, no waiting for it to finish.
NovaThe cheapest is a one-stop at 10:15 for $189, then a nonstop
YouOnly nonstop, please.
NovaGot it, nonstop only. The 7:40 is best, $206 each.
It asks before it pays
Three steps always wait for your Allow:
PayingPlacing a callSharing a file
PayPay $808 for 2 flights and 3 nights with Visa ending 42?
DenyAllow
Allowed · paid and booked
It asks back
When something is missing, one short question instead of a guess.
YouPlan Bangkok for two in December.
NovaSure. Which dates, and what budget?
YouThe 11th to the 14th, about $1,500.
Your language
Say “speak Japanese” and it answers in Japanese from then on. Here is its own reply in each.
日本語わかりました。日本語でお話しします。
A voice you like
Pick how Nova sounds, or say “speak like a kid”.
The mascot's ownA womanA manA kidAn animalOpenAI's natural voicesYour computer's
Three jobs at once
Hand over a job and keep talking; each tells you when it is done.
Compare three suppliersSI agent
Fill in the visa formSI agent
Summarize forty reviewsSI agent
It remembers the conversation
Each day's voice conversation is saved as a chat in History, so you can ask about it later, or turn that off.
History · yesterday's voice chat
YouWhat did we decide yesterday?
NovaThursday at three for the launch plan, with Sarah, Ken and Priya, on a video call.
The film
Six scenes. Pick one.
Each plays in the film at the top, with the music and both voices. Names, numbers and prices in them are made up.
An example day
A day with Nova.
Eight things said out loud between breakfast and bed. Nova does the clicking; you do the deciding.
Read me my brief.
Today's meetings and the new notifications, read aloud while the coffee brews.
Send Daniel this month's invoices, in one sheet.
Twelve PDFs read into one Excel sheet with a total, shared on WhatsApp after your Allow.
Summarize this page.
A long report becomes its main points, told to you while you keep reading something else.
Book a table for four near the office at one.
Places nearby checked, one booked, the confirmation in your inbox.
Set up the launch meeting with Sarah, Ken and Priya.
A free hour found, the invite sent with a Meet link, everyone told.
Call Ahmad.
It asks first, places the call and leaves the talking to you, then sends him the address.
Plan Bangkok for two, and book it.
Flights, a river-view hotel and dinner, $808, paid only after your Allow.
Play rain sounds, and stop in 30 minutes.
It plays them on YouTube and closes the tab at 23:30.
Say it
The way you would say it.
One sentence each. Nova works out the steps.
Summarize this page.
Play lo-fi on YouTube, and close it in 30 minutes.
Every morning at seven, read me my brief.
Find three suppliers and compare their prices.
Watch this price and tell me when it drops.
What did we decide yesterday?
Open my notes.
Switch to the next tab.
Reply to Sarah: Thursday works.
Book a table for four near the office.
Fill in this form with my details.
Find the cheapest flight to Tokyo in March.
Turn this sheet into a chart.
Send Daniel the October invoices.
Who messaged me today?
Speak Chinese with me.
Call Ahmad on WhatsApp.
Add the launch plan to my calendar.
Read me the comments on this video.
Download every PDF on this page.
Close the other tabs.
Speak like a kid.
Put this month's receipts in one sheet.
Stop.
The details
What to know before you talk.
Its name
Say “Minda” before a request. Or turn on Listen without “Minda” in the microphone menu: anything that sounds like a request is then acted on. It is off by default.
Where your voice goes
Each stretch of speech, a 16 kHz clip of at most 12 seconds, is sent for transcription: to OpenAI with your own key, or to Minda's hosted models on your credits. A clip starts only when a human voice is heard. With Keep SI on this computer, Windows' own recognizer hears you instead, in US English.
When it asks
Before it pays, places a call or shares a file, the step waits in the chat with Allow and Deny. You do the talking on a call.
Interrupt it
Talk over a reply and it stops at once and takes what you said as the next request. “Stop” ends a command that is taking too long.
Voices and languages
The mascot's own voice, a woman, man, kid or animal voice, one of OpenAI's natural voices, or your computer's. English, detect the language, or one of 20 languages.
Saved
Each day's voice conversation is a chat in History, searchable and synced like your other chats, unless you turn that off.
Cost
Every plan, Free included. Transcription and natural voices use your own OpenAI key, which costs no credits, or Minda SI credits as you use them: Free has 100 a month, Pro 2,000. Your computer's own voice costs nothing.
Nova
One of the agent's eleven looks: a swarm of light that takes the shape of each task and stays light on any computer. Pick it in the agent's menu.
Say “Minda”.
Then say what you need. Nova takes it from there, and checks with you before anything costs money.