The assistant layer for RistOS devices
Press the button on your wrist and say it out loud. Rist works out what you meant, calls whichever tools the request needs, and hands you back the finished thing.
It knows the time and where you are. Live transcripts are held two minutes, then discarded.
You said
“Text Maya I’m running ten minutes late.”
Rist
Ready to send — one message, drafted
“Hey — running about ten minutes behind. See you soon.”
“Want me to send it?” Nothing leaves your wrist until you say so.
Two minutes after this exchange ends, the transcript of it no longer exists.
What you can ask
No wake-word gymnastics, no command syntax to memorise. Because it already holds the clock and your location, the vague parts of a sentence — later, home, my next one — resolve on their own.
“Remind me to grab the dry cleaning when I get home.”
Reminder saved and tied to a place, not a clock. It fires when you actually arrive home — even if that’s Thursday.
RemindersLocationDone instantly
“How long until my next meeting?”
Spoken back: “Twenty-two minutes. Design review at 2:30 with Priya, in the Cascade room.”
CalendarDone instantly
“Navigate to the trailhead, avoid the highway.”
Turn-by-turn started. 41 minutes on surface roads, spoken one turn at a time so you never look down.
MapsNavigationDone instantly
“Read me the last email from my landlord, then reply that Saturday works.”
Reads it aloud, drafts the reply, then reads the draft back and waits for a yes before anything sends.
EmailConfirms first
“Add oat milk and limes to the shopping list, and set a timer for the rice.”
Two things, one sentence. List is now seven items; timer running for eighteen minutes.
Shopping listTimersDone instantly
“Is it going to rain before I finish this walk? Put my podcast back on after.”
“Not before six — you’ve got about two hours.” Then the episode resumes exactly where it stopped.
WeatherPodcastsDone instantly
Your call, not our policy
How often it stops to ask is a setting, and the setting is yours. Three levels, from acts-first to nothing-happens-without-a-yes. It arrives on the middle one, and you can move it whenever you like — say “stop asking me about texts” out loud, or change it in Settings. Pick a level below and watch what it does to the rest of the panel.
The side button is the microphone. Let go and it stops — there is no always-on listening.
Whatever level you land on, the confirmation itself is the same three beats. On Hands off you skip straight past the middle one for almost everything; on Ask every time you get all three, every time.
STEP 01
One press, one sentence. The button is the only way it starts listening — there is no open microphone and no wake word waiting in the background.
STEP 02
When your level calls for a confirmation, you get the actual recipient and the actual words — not a summary. If it picked the wrong Maya, you hear that before it matters.
STEP 03
“Send it” sends it. Anything else — “change that,” silence, a lowered wrist — leaves the draft sitting there, unsent.
How it routes
Rist is the LLM orchestration backend behind the button. A router reads your sentence alongside the injected time and place, works out which capabilities it touches, calls them in an order that makes sense, and returns one finished answer — not a list of results.
You say
“Text Priya I’m running late, and start navigation home.”
transcript window 02:00
Ambient context, injected
Router
One request → three capabilities
plan · call · merge
Capabilities called
Confirm
“Send to Priya: running about fifteen late?”
you say yes → it sends
Resolved — 1.9s
Done
Priya notified · 6:42 PM
Navigation started
Home · 18 min · I-5 N
Anatomy of one request
Most assistants stop at the first tool call. The interesting requests are the ones that need three or four, in order, where the output of one becomes the input to the next — and where getting it wrong sends a message to the wrong person.
You, at 8:04 AM
Move my 3pm to tomorrow and tell Priya.
Nine words. No meeting name, no date, no phone number, no time zone. Everything else has to be inferred, looked up, or asked.
Execution trace
context.inject() 0 ms
Every request arrives pre-loaded with the wall clock and your location. Thu 8:04 AM, Seattle. That is what makes “3pm” and “tomorrow” resolvable at all.
calendar.find(day: today, near: 15:00) 180 ms
One match — Design review, 3:00–3:45 PM, with Priya Raman. If there had been two, it would have asked which one rather than guessing.
calendar.move(to: Fri 15:00) 240 ms
“Tomorrow” resolves against the injected clock, not the model’s guess. Time is stored as an absolute instant, so a flight across time zones doesn’t move your meeting.
contacts.resolve(“Priya”) 90 ms
Exactly one Priya, and she’s already on the invite — two independent signals agreeing. Ambiguity is a question, never a coin flip.
messages.draft(to: Priya) 310 ms
Drafted, not sent: “Hey — moving our 3pm to tomorrow, same time. Works?”
confirm.await(voice) held
It reads the message back and waits. Nothing leaves the device until you say yes. Every outbound action is gated this way — messages, email, calendar invites, items added on someone else’s behalf.
messages.send() 200 ms
Sent. Total wall time from button release to spoken confirmation: 1.02 s, excluding the pause where it waited on you.
“Moved to Friday at three. Priya’s been told.”
Spoken back on the wrist. Your eyes never left the road.
The capability field
There are no apps to open and no menus to learn. Each capability is a typed tool the router can call, alone or in combination. Say the thing you want; the composition is our problem, not yours.
New capabilities are added to the field over time and land on every plan they belong to — nothing to install, nothing to update on the watch.
Wednesday
Nothing here is a demo. It is a Wednesday — an alarm, a commute, a text you meant to send, milk you almost forgot. Watch how little of it you have to think about.
Each hour below carries its own light. The page warms toward midday and cools into the evening with you.
6:40 amStill dark out
“What does my day look like?”
The alarm fades and you ask before your feet hit the floor. Three things, in order, spoken quietly back to you — standup at nine, the doctor at half eleven, a review after lunch. Then one more line you didn’t ask for: rain arrives around four. It knew that mattered because it knows you walk home.
8:15 amFront step
“Navigate to the office.”
No address. No app. It knows where you’re standing and which office you mean, and the first turn is already on your wrist before the door clicks shut. Turn-by-turn, one instruction at a time, a small buzz at every corner. Your phone stays in your pocket the whole way.
8:20 amThird block, walking
You didn’t say anything at all.
No button, no wake word. The commute started, so the episode you abandoned yesterday picked up at eighteen minutes and thirty-two seconds, exactly where you left it, ducking politely under each navigation prompt and coming back after. This is the one moment in the day where the right answer was to not be asked.
11:00 amBetween meetings
“Remind me to call the plumber when I get home.”
Not at six. Not in four hours. When you get home — a place, not a time, which is how people actually think. It confirms in four words and then does nothing at all for seven hours, which is the hard part.
1:30 pmHalf a block from lunch
“Text Maya I’m running ten minutes late.”
It finds Maya, writes the message, and then — before anything leaves your wrist — reads it back and waits for a yes. Nothing is sent on a guess. You say yes without breaking stride, and the little buzz that follows is the message going out.
4:45 pmHands full, kettle on
“Add oat milk to the shopping list.”
Two seconds, no unlock, no scrolling to find the right list. It’s the smallest thing the assistant does all day and probably the one you’d miss most — because the alternative is remembering it in the shop, on Saturday, alone.
6:10 pmKey in the door
“Call the plumber.”
Seven hours later, said back to you at the exact moment it’s useful. Not a notification you’d swipe away at a red light — a reminder that waited for the doormat. This is the whole idea of the product, and it takes four seconds to happen.
9:30 pmLights off
“Keep reading the book.”
Chapter twelve, from the sentence you fell asleep on. The screen goes dark within seconds and stays dark, because nothing about this needs to be looked at. Somewhere in there you stop listening. It notices, and stops too.
What makes it work
Most assistants start from zero every time you speak, which is why you end up talking like a search box. Rist begins each request already holding the time and where you are — injected before the model sees a single word. That is the entire trick, and it is why a whole day of ordinary sentences just works.
Not just the clock — your clock. What has already happened today, what is next, how long you have before it starts, and whether “tonight” means in four hours or right now.
“How long until my next meeting?”
Here, home, the office, the route you are on and the one you are headed down. Enough that you can say a place instead of an address, and mean a moment instead of a time.
“Remind me when I get home.”
Reading a message back, naming the contact, saying the recipient out loud. The gate lives in the orchestration layer rather than in the model’s instructions, so no phrasing talks it into skipping the step.
“Send to Maya Ellis — yes or no?”
Quiet by design
That is how long a live voice transcript exists. Long enough to understand you and act on it. Then it is discarded — deliberately, not incidentally.
Illustrative. The chart is a drawing of one spoken request and the two-minute window it lives inside — not a recording of anything.
0:00
0:01
0:03
2:00
Two minutes is deliberate: long enough for a natural follow-up, short enough that there is no history to leak or sell. It is a fixed ceiling, not a default we plan to widen.
Every assistant is shaped by what pays for it. We can’t tell you what another company does with what it hears — we can tell you what ours is structurally rewarded for.
When revenue comes from attention or placement, more time spent and more signal collected are both worth something. That doesn’t make a company dishonest. It does mean the product and the person can want different things, quietly, forever.
A flat monthly fee is the only thing on our side of the ledger. We don’t earn more when you use the device more, and we have nothing to sell that gets better with a longer record of you. The best outcome for us is that it works, quickly, and you put your wrist down.
Because you asked for it, or because the service can’t run without it.
Not stored, not built, not part of the product.
What we’re not claiming: we’re not going to tell you your data is untouchable, or wave a certificate at you. What we can tell you is the retention window, the confirmation step, and how the company makes money. Those three facts are the whole argument.
Plans
A plan decides which of the twelve tools the router is allowed to call on your behalf. Requests aren’t metered and there’s nothing to run out of — you’re choosing how much of the field you want switched on.
$25/ month
The assistant most people actually use all day: time, people, memory and the next thing you have to do.
Included7 / 12
$35/ month
Core, plus the tools that matter once you leave the house — turn-by-turn navigation, email and the list you keep forgetting.
Included10 / 12
$50/ month
The whole field switched on, including the two tools that reach past your own data and out to the world.
Included12 / 12
Pricing is provisional. These are the tiers we intend to launch with while the first production run is finished. They may move before general availability, and anyone who signs up early will be told before anything changes.
What a plan buys. A tier decides which capabilities the assistant is allowed to use on your behalf — nothing is metered and there is nothing to run out of. Requires a Rist device, sold separately. Same two-minute window and the same approval controls on every tier. Change tier or cancel in any month.
Questions
Still unsure about something? Ask a person at hello@ristmobile.com.
The assistant layer itself — the LLM orchestration backend that runs behind RistOS devices. Your monthly price buys a set of capabilities the assistant can use on your behalf: which tools it may reach for, and on whose say-so. Requests aren’t metered, and there is nothing to run out of.
When you speak, one request goes up. A router reads it alongside the injected time and location, works out which of the twelve capabilities it needs, calls them in an order that makes sense, feeds the output of one into the next where required, and returns a single finished answer.
The alternative — the thing most assistants do — is match your sentence to one tool and hope it was the right one. Requests like “move my 3pm and tell Priya” need three tools and a lookup between them.
Messages, email, contacts, calendar, notes, reminders, timers and alarms, maps and turn-by-turn navigation, weather, your shopping list, podcasts and audiobooks, and web search — all by voice, all with the time and your location already in hand.
Audio is captured only while you hold the button. It is transcribed, the assistant reasons over the text and uses whatever tools the request needs, and the live transcript is held for two minutes so follow-ups make sense — then discarded. Audio is never written to disk on our side, and there is no transcript history for you or us to browse.
What persists is the output you asked for: the message you sent, the reminder you created, the note you dictated. You can delete those by asking.
Because the request has to finish. Two minutes is long enough for the assistant to understand you, run the work, read the result back and take a follow-up — and short enough that there is no meaningful record sitting around afterwards. It’s a deliberate number, not a default, and there is no longer tier and no toggle to extend it.
No. There is no wake word and no ambient listening. The side button is a physical gate on the microphone: press to open it, release to close it. If your finger isn’t on the button, nothing is being captured — and there is no feed or background app waiting for you to come back.
No. Anything that changes the world outside the watch — a message, an email reply, a calendar change that other people are on, a deletion — is drafted, read back to you with the recipient and the wording, and held until you say yes. The gate lives in the orchestration layer rather than in the model’s instructions, so there is no phrasing that talks it into skipping the step.
Reads are not gated. Checking the weather, your next meeting or a route just happens, because there is nothing to undo.
It asks. Ambiguity — two contacts with the same first name, two meetings near 3pm, two Kims on the invite — produces a one-line question rather than a guess. The read-back before sending is the second net under that.
Every request arrives with the current time and your current location already attached, before the model reads it. That is what lets a relative phrase — home, tomorrow, in twenty minutes, when I get there — resolve to something concrete. “Remind me when I get home” becomes an arrival trigger; “is it going to rain” means where you are standing. The location is used for the request in front of it and is not retained as a history.
Almost never. The watch face shows a glanceable confirmation — the next turn, the drafted message, the item that was added — but every one of those is also spoken. There are no apps to open, no notifications stacking up and no feed.
Anything already running on the device keeps running — timers, alarms and a navigation route that has already been fetched. New requests that need the assistant will queue and complete once you have a connection again, and it will tell you plainly when it can’t reach the service rather than pretending it did the thing.
The subscription. That’s it. There’s no advertising product, no placement to sell inside the assistant, and no revenue line that grows when you use the device more. If the assistant answers you in eight seconds instead of forty, that’s a better outcome for us too.
No, and we’d rather be plain about it than sell you a sentence we can’t back. We’re not making claims about encryption schemes or certifications on this page. What we’re telling you is narrower and checkable: the retention window is two minutes, sends require your confirmation, and there is no ad business here.
Any RistOS device. Today that means the Rist watch; the assistant layer is designed to be device-shaped rather than watch-shaped, and the Full tier lets one assistant follow you across more than one of them.
Change tiers any month, in either direction; it takes effect on the next cycle. Capabilities you drop go dormant rather than being deleted, so the notes, events and lists you already made are still there if you come back.
Cancel and the assistant stops — that is the part you are subscribing to. Your notes, events and lists remain yours and exportable, and the two-minute window means there is no back catalogue of your voice to worry about either way.
It is provisional. $25, $35 and $50 are the tiers we intend to launch with, and we would rather show you the shape of the thing now than a blank space. If it moves before general availability, early sign-ups hear about it first.
Press the button, say the sentence, get on with your day. One button, twelve capabilities, and an orchestration layer whose whole job is working out which of them your sentence actually needed.
Requires a Rist device · pricing provisional · transcripts held 2 minutes