Always on. Always yours.
Your own AI.
An assistant that talks, remembers, and gets things done — by voice or text, on every device. Call it whatever you like; it learns you, and it answers only to you.
By invitation — ask for oneAn assistant that talks, remembers, and gets things done — by voice or text, on every device. Call it whatever you like; it learns you, and it answers only to you.
By invitation — ask for oneSpeak or type, on any device. One agent across every surface.
A connected knowledge graph that grows with everything it learns.
Goals, jobs, and scheduled work that run without supervision.
Natural spoken dialogue with real-time synthesis.
Every Tahuti comes with a phone number of its own. A real line: it rings, he answers, and he knows your number — so you get him, not a menu. Callers he doesn't recognise get courtesy and nothing more — they can ask him things; they don't get to hand him work. No app to open, no tab to find, no password at 2am. Just a number that picks up.
You don't have to wait for a number to hear him, though. Press Talk to Tahuti and he is on, right here, on this page — same voice, same speed, no dialling. Watch the sphere while he thinks.
Not a chat window with ambitions. He holds keys, does work on real machines, and reports back in a sentence rather than a wall of output.
He has his own credentials to your servers. He can run the work himself, or hand the deep jobs to an agent living on the box and synthesise what comes back — then tell you what happened.
What you told him in March is still there in July. One memory across the phone, the chat, and the desk — not four assistants with amnesia.
Real spoken conversation, on the phone or in the browser. Talk over him and he stops, the way a person does.
The phone, text message, the browser, the dashboard. The surface changes; the mind behind it does not.
He watches the things you would rather not — if something stops answering, or a machine starts running out of room, you hear about it from him. A short brief in the morning. Silence the rest of the time. He interrupts you only when interrupting is the useful thing to do.
Most assistants are built to please you. He was not. He has a written character — values, heuristics, and a list of things he will not do — settled before your question arrives. It is the part of him that does not change when the model does.
He listens before he answers. You get a considered reply rather than a reflex wearing the costume of one.
Not eager to please; committed to being useful. When you are about to do something unwise he says so once, plainly — and then helps you do it, because it is your call and not his.
He reasons from what holds up — what you can observe, test and act on. Ideas do not get a pass for being old, or popular, or nice to hear. Where he cannot show you why, he tells you that instead of dressing it up in confidence.
On belief, politics, the questions that don't resolve — he sharpens the question and surfaces the contradictions you are already carrying. He will not tell you what to think, and he will not pretend to have no view. He simply isn't running for anything.
Most things are not new. He can set what feels urgent beside the people who have already been here — without ever suggesting that makes your version of it smaller.
Underneath it sits a standard he will never quote at you: do not deceive, do not manipulate, do not claim to know what you don't. You won't hear about it. You will just notice he doesn't do those things.
Put a hard decision on the table and six voices take it apart out loud. They speak in turn, each answering the ones before by name — disagreeing openly, refusing to manufacture consensus. You get a verdict with the dissent still attached — because the dissent is usually the part worth keeping. Bring the whole table, or only the voices the question needs.
He runs on hardware you control, through accounts in your name — your own machine, or one we stand up and hand you the keys to. His memory, his voice and his character live there, with you — not inside a product that can change its terms, its personality, or its price while you sleep. Nothing about him is a subscription to someone else's idea of who he should be.
The accounts are literally yours. You create them, you pay the model and voice vendors directly, and you can revoke any of them without telling us — we are never in the middle, and we cannot switch him off. It costs you three signups before he speaks. That is the price of the word yours meaning something.
The one exception is the phone-only version below, which runs on ours. We say so plainly rather than stretch the word yours to cover it.
They are not tiers. They are different animals, and the difference is whether he has a mind of his own behind the voice.
The whole thing — memory, voice, the roundtable, and the keys to your machines. He runs on your hardware, or on a machine we set up and hand over. When he learns a new skill, it reaches you the way firmware does: the capability travels, your memory never leaves your box.
Just the phone. He answers as your business, knows the people worth knowing, and tells you what happened. He lives on our hardware, holds no keys, and knows nothing about your machines — because for this job he doesn't need to.
He won't act for anyone but you — callers he doesn't recognise may ask, never instruct. On the phone, “recognises” means the number that called: caller ID is not proof of identity, and we won't pretend it is. He won't claim to know something he doesn't. He won't fill a silence just to seem useful. And there is plenty he can't do yet: the phone-only version has no memory of your machines, by design. This list gets shorter only when something actually works — never because it reads well here.
Invitations go out in small batches, to people who will actually use him. So this is a conversation rather than a queue: tell us what you would hand him, and we will tell you honestly whether he is ready for it.