How hiy is set up

Making a twin or an agent is a guided create flow — six screens: the first five ask, and the last one shows the work happening. There is one typed field in the whole of it. Nothing blocks you: every question has a sensible default or a skip, and nothing at all is written until you press the button on the last screen that asks. Walk away before that and no agent exists anywhere, so an abandoned flow costs you nothing — not an agent, not one of your slots, not a half-made thing to go and tidy up. It’s built for a phone as much as a desktop — every screen was measured, not assumed.

Back is on every screen but the first and the last, and it keeps what you’ve entered — in that tab. Step back to check the address and everything you typed is still there when you come forward again. Reload the page and you land back at the first screen with what you typed still filled in — the answers are kept in that tab, and only that tab. Nothing is sent anywhere, closing the tab forgets them, and how far you had got is never stored: you walk forward through screens you have already answered rather than resuming a half-finished form. Back never undoes anything either, and up to that one press there is nothing to undo: no agent, no import, nothing kept. After it there is no Back at all — the last screen is the work, and a button walking you back onto a press that has already been spent is a button that cannot fire. For a different kind, make a second agent.

A ✕ closes the whole thing, and it is on every screen. Top right, labelled Close and go back, and that is what it does: it puts you back on your dashboard, with the agent you opened this from already the one selected if there was one. Back walks you one screen; this leaves. If you have entered anything at all it asks first — Leave without creating this agent? — offering Stay on this page and Discard and leave, and Discard is the plain word for it: nothing you entered is kept, and coming back later starts from the top. It asks nothing when there is nothing to lose, because a confirmation over a form you have not touched is what teaches people to click through the ones that matter. Screen 1’s Back to your dashboard button is still there too; the ✕ is that exit made available from every other screen as well. The last is the one place it behaves differently, because by then your agent exists — see what happens when you press Create.

They are answers, not settings you have to keep up to date. Nothing is copied or frozen: your dashboard, your empty screens and which surfaces appear are worked out fresh from your answers each time a page loads. Change one and the next page you open reflects it.

The six screens#

Six for a public twin; five for a Support Agent or a private twin, which skip Tools — screen 4 below says why. The marks under the logo are named — Kind, Address, Pages, Tools, Channels, Build — and the one you are on is filled. Wherever the agent is doing something — checking your address, finding pages, reading them — hiy’s mark moves beside the words; while it waits on you, the mark is still.

1. What are you creating?#

Two cards: A twin or A Support Agent. Choose the twin and one more question appears — Who can reach it?, answered Public or Private. Public is chosen for you.

If you already have a public twin, or your three agent slots are used, or your Founding trial is spent, this screen says so before you type anything — as a badge on the card it concerns (2 of 3 slots) and a sentence beneath, with the option that can’t succeed dimmed rather than hidden. A Support Agent and Private both say what they cost, under their own card at all times: either one starts your Founding trial, if you still have one, and uses one of the three agent slots.

While Early Access spots are open, there is one offer instead of two: a Support Agent or Private made here starts Early Access rather than the Founding trial — free for a month, on limits of its own, with one agent slot. The cards say so, and a short note says what Early Access is and what happens when it ends; the Founding trial isn’t mentioned. If your trial is already spent, the note offers Claim your spot instead.

That toggle is the whole difference between two of the three kinds this documentation describes: a public twin is one anyone can reach at your hiy link, and a private one is what the rest of these pages call a Team Agent — the same twin, answering only the colleagues you invite. See The three kinds of twin for what each is for.

Three answers here cannot be changed later from a settings row — the kind, the tools and the channels. The kind decides the voice, how long answers may be, what the twin may bring up, and who can reach it at all; if you want a different kind, make a second twin, you can have more than one.

Choosing a Support Agent reveals its presets underneath — a small set of named setups, plus Start from scratch — on the same screen, in place of the audience toggle. The first preset is preselected; every preset shows the scope it declares, and the chosen one also shows the first words of the opening line that bundle writes — opens with: “Hi — I answer from our product docs…” — so you’re reading a real choice rather than guessing; one click changes it. See Four presets, one agent.

Beside the form, on every screen, is a preview of the agent you are making: its avatar and name (the kind, until it has a name), a status line saying what the agent is doing right now — Waiting for your answers, then Looking for the pages under acme.com…, then how the import is getting on — and, once it has one, the opening line it will greet with, shown as a single bubble under How it opens — a preset’s line from the moment you pick one, and the agent’s own once it has written itself an introduction on the last screen. The address, what it covers, the tools you noted and where it will live are listed beneath. On a phone there is no beside: the same preview sits at the end of the screen under Your agent so far, folded until you open it — and it opens itself from the moment there is an import to report, because that is the one line on it that changes while you are reading it. If you opened this screen from an existing twin’s dashboard, a Back to your dashboard button returns you to it. A brand-new account never sees that button — there’s nowhere yet to go back to.

2. One address#

One field, worded for what you picked — your docs, your internal material, or something of yours we can read. Two fields under it fill in from what you paste, both yours to change before you continue:

• Address — shown as Your hiy address for your own twin, or plainly Address for a Support or Team Agent, whose address never becomes a public link. Worked out from the address you pasted (docs.acme.com becomes acme) and checked live as you type: while the check runs the agent’s mark moves beside Checking acme…, and the answer lands as a badge — acme is yours · available, or the reason it is taken. A taken or reserved candidate is retried once with a suffix, and then left for you to edit.
• Name — the same label, title-cased.

Both carry a small suggested mark until you edit them, so you can tell what was worked out for you from what you typed.

Arriving with your site. A link that already names it (hiy.ai/app?variant=support&url=acme.com) keeps it through sign-up, and it lands in this field with the two under it filled in, exactly as if you had pasted it. Nothing is read until you press Find its pages.

An address behind a sign-in. If your handbook or wiki is only reachable once you have logged in, nothing can be read from it — we arrive as an ordinary reader with no account, and the site refuses every page. The next screen says so in those words rather than reporting a missing sitemap, which is what it used to do and sent people hunting for one. Export the material and add the files from Agent knowledge instead; the twin is still created, and nothing is lost. This is also why the internal wording asks for a handbook or an internal site and not for an export: this field takes an address, and an export is a file.

Not on a site? A row of chips beside it records where else your material lives — Files on my computer, ChatGPT, Claude or Gemini, Lists of things, plus Notion — Soon. What you tick is noted against your twin, so your agent’s Overview knows what you still mean to bring. A blank address is allowed too: the twin is created with nothing to import yet, and the receipt says so plainly rather than pretending an import is running.

Then one tick in its own bordered row, confirming this represents you or something you own, and Find its pages — it reads Continue while the address is still blank — which starts looking at your address. Nothing is made here: the press lists the pages it can find, and reads none of them. That’s the next screen. The press that cannot be undone is the last one, on screen 5 — so up to that point you can leave, or go back and change any answer, and no agent exists anywhere.

Why the tick is here and not beside the button that makes the agent. It is a declaration rather than an answer, and this is the screen where it starts to matter: pressing Find its pages is the first moment anything of yours is looked at, so the tick sits immediately above the address it covers. It is asked once. The last screen names it back at you, as the last row of the card above the press — You confirmed acme.com represents you — so the one irreversible act in the flow isn’t silent about the claim that licenses it. That line is a reminder, not a second tick: there is nothing extra to agree to and nothing extra to press.

The button is never greyed out. Press it with something still missing and it says which one thing — the address, the name, the tick — and puts your cursor in that field, and the sentence goes as soon as you supply that thing. A greyed button that names none of the four things holding it back is the same information withheld.

Step Back onto this screen later and it is the same form, with everything you entered still in it. It made nothing, so there is nothing for it to report back at you, and pressing Find its pages again simply looks at the address again — change the address first and it looks at the new one. Your address on hiy is still a once-only decision, but the moment it is fixed is the press on screen 6, not this one — see The address cannot be renamed.

3. Which pages should it read?#

While the pages are being found the screen shows the agent looking — Looking for the pages under acme.com…, its mark moving — and the result is stated the same way, settled: Found 3 pages under acme.com. Under it, the list of everything found, before a single page has been read. Sites have more pages than anyone wants an agent to learn, and picking twenty good ones beats importing four hundred.

Pages are grouped by their first path segment — /blog, /docs, /pricing — with a count beside each and a checkbox at both levels. Sections of a single page are folded into Top level rather than given a heading of their own. Everything starts ticked, so the default is your whole site and you take away what you don’t want — unless the address you typed has a path: give acme.com/docs and only the pages under /docs start ticked, with the rest of the site still listed one tick away. The room your plan leaves is counted on the plan this agent will have: making a Support Agent, a Team Agent or a private twin on a Free account starts your trial, so it is the trial’s room you see. There’s a filter box for finding things by path or title, a running N of M selected with a meter beside it in the docked bar, and Add a link we missed if something isn’t in your sitemap — one address per line, or several pasted at once.

Above the list, tags say what was seen in the groups that look like duplicates of what’s already there — 12 groups look like a language version, 3 look like an archive listing, 8 numbered list pages — and one button unticks all of them at once. Three things are looked for, and each says what it actually saw rather than announcing a verdict:

• A language version. A first segment that is a language code, with a count of how many of its pages repeat a page found elsewhere in the list — “es is a language code, and 38 of these 40 paths repeat a page found elsewhere in this list.”
• An archive listing. /tag, /category, /author, /archive and the like: pages that re-list posts you already have.
• A numbered list page. /blog/page/2, ?page=3. This one is judged page by page, so a numbered listing never condemns the real posts beside it.

None of this is AI. No model runs at this point; these are three tests against the paths themselves, which is why each one tells you what it matched instead of asking you to trust a badge. Disagree and untick nothing — the flags change no default, they only make one press available.

Sizes aren’t shown, and that’s deliberate: nothing has been fetched yet, so nobody knows how long any of these pages are. You’ll see word counts under Knowledge once they’re in.

Import these N rides in a bar docked to the bottom of the screen, with Skip for now and Back on one row beneath it, so all three stay reachable without scrolling the list first. Import these N notes exactly what you ticked and moves on — the reading itself begins when you press Create on the last screen, because there is no agent to read into yet. Skip for now imports nothing, and the receipt then tells you what you left — “Nothing imported — 87 pages were found. Back to pick some, or add them later from Agent knowledge.” The count appears even if you skipped before the list had finished loading; nothing reopens the choice for you, Back is how you return to it. Where there was nothing to pick from — no address, or an address with nothing under it — that same press reads Continue without importing.

A YouTube channel gets the same screen as a flat list of video titles — a path tree would say nothing about it.

If your plan has room for fewer sources than you’ve picked — for a first Support or Team Agent, the plan it starts: Early Access or the trial — the room is stated in that same bar, directly above the button: “Room for 5 more on your plan — the import stops there. Untick a few, or free room up or raise the limit.” It’s a fact rather than a gate: it unticks nothing for you and refuses no tick of yours. Carry on and the import stops at the first source past your plan’s room and names it; everything behind that one is left untouched, and it goes in once you free room up or raise the limit — until then, adding a source anywhere is refused with your plan’s limit named. If the list itself can’t be fetched, the reason appears right here and Back is how you try a different address — nothing was sent, because nothing is sent from this screen.

Step Back here later and the list is exactly as you left it, ticked the same way, with any links you pasted still in it. Nothing has gone anywhere: the pages you pick travel with the press on screen 5, and they are sent once.

4. Which tools do you use?#

One card, What your agent will be able to do, and one thing on it works today, only for a public twin: a booking link to Cal.com or Calendly. Those two are tick rows with a Works mark, and ticking either shows a line under the row saying what it still needs from you — the booking link itself, which you add later under Lists.

Everything else is noted, not wired, so it does not crowd out the two that are: the rest fold under one line in the same card reading “10 more on the way”. Open it to see them — Stripe, Shopify, Slack and Twilio under Actions; Zendesk, Intercom, HubSpot, Freshdesk, Help Scout and Salesforce under Helpdesk — each carrying a Soon mark, and none of them clickable. The whole screen still skips in one tap.

A Support Agent and a private twin do not see this screen at all. Neither kind may present what you offer, so a booking link is removed from their answers before a visitor ever sees one — Cal.com and Calendly included — and every other connector is on the way. A screen with nothing to answer is not a step, so their flow runs five screens and the marks under the logo name Kind, Address, Pages, Channels, Build.

5. Where will it live?#

Channel rows, filtered to what your kind can actually use, and different for each: a public twin sees live chat on your site, inline on a page, and your page on hiy.ai; a Support Agent sees live chat, inline, and inside your product by token; a private twin sees only your colleagues, signed in. All of those are live today. WhatsApp, Slack, Email and Phone are named under On the way on every kind, not wired.

Each row says what it needs before you choose it — that “inside your product” wants your server to mint a token, that your page on hiy.ai is live the moment you publish — and the two embeds, live chat and inline, carry a one paste mark, because Publish hands you one line for each.

This is the last screen that asks, and its press is the one that makes the agent. It names the act rather than saying “Continue” — Create my twin for your own twin, public or private, and Create Support Agent for a Support Agent — because everything you have answered is written in that one moment, and the pages you picked on screen 3 begin to be read. There is no Skip here: a second button beside it would create your agent with one of your answers thrown away. Back still works, right up until you press it.

Above the button, a card — When you press Create — lists what the press will do, from what you have answered: make the agent, read the pages you picked from your address (or that it reads nothing yet), take its icon and colour from the site, and wait for you before it trains. Its last row names the address you ticked on screen 2 back to you — You confirmed acme.com represents you — because the tick was screens ago and this is the press it licenses; it asks nothing of you, and it isn’t shown when you gave no address.

There used to be a different sixth screen that opened a chat with your new twin. It went on 2026-09-04, because it was lying: the flow reaches that point about ninety seconds after you start, and reading and indexing a real site takes minutes to tens of minutes. The chat was mounted under the words “sources are in and indexed” while the indexing had not begun. Your Overview says what has genuinely happened instead, and the Sandbox is there the moment there is something to test.

6. What happens when you press Create#

The press is the only moment anything is written, so the screen after it is the only one that has work to show. Its heading follows the work: Your agent is on its way, then Your agent answers now once your pages answer, then Your agent is ready when its summary is written. It shows four things, in order, and each one moves because the job said so — never on a timer.

Making your agent. The single request that carries every answer you gave. Until it comes back there is no agent, no draft and no half-made row anywhere; if it is refused — a plan limit, a handle already taken — you are put back on screen 5 with the reason and the button live again.

A photo and a colour arrive from your own site, just after that press. When you gave an address, hiy looks at that page for the icon it declares — an apple-touch-icon or a rel="icon" — and for its theme-color, and uses them as your agent’s photo and accent. That look-up needs an agent to write onto, so it happens the moment there is one, and this screen says where they came from — a badge on the Making your agent row, from acme.com, reading Icon and colour taken from acme.com. To change either, go to your agent’s Profile tab — Photo and Accent colour are both there, and the create flow no longer asks. Two things the look-up will not do: it maps a declared colour to the nearest of the six accents rather than inventing a seventh, and it refuses greys, whites and blacks outright — those are usually the colour of a browser toolbar, not of a brand. A site with no icon and no declared colour changes nothing, and nothing ever waits on it.

Reading your pages. The pages you picked on screen 3 are fetched now, and the count is the job’s own, said once, in the header of the stream beneath: Reading acme.com · 12 of 60 · 0:42. The step row itself keeps its label, its state and a bar. The count moves in jumps rather than smoothly, because the importer reports at the end of each batch — a number that sits still for half a minute is a healthy job, not a stuck one. If you picked nothing, this step says so and nothing is claimed for it.

Under that count sit the pages themselves, one line each as the job walks them: the one in hand marked reading, the ones behind it read, and anything that could not be fetched failed with the reason under it — 403 from the site is the difference between a page to try later and a page that is theirs to open — newest at the foot with the older lines scrolling up out of the box, which keeps one size while the read runs and ends at its last page once it is done. Above it one line says what the agent is doing — Reading acme.com · 4 of 8 — with a clock beside it while work is running, and the stream closes with the tally: All read · 7 pages in and 1 failed. When the importer reports a whole batch at once the cards are revealed one after another rather than all together, so a batch reads as a stream and not as a refresh. It is the same report the count is drawn from, said in names instead of in a number — so the jump that leaves the figure motionless arrives as pages you can actually see being taken in. The clock is the one thing here that is not the job’s report: it is how long this tab has watched the read, there to tell a slow page apart from a dead job. No line changes because of it.

Its own introduction. When the read has settled, your agent proposes how it would introduce itself: a greeting — the first thing a visitor hears — and an About line, what this is in a sentence or two. Both are drafted from the pages it just read, and the card says which (From the 8 pages I read); with no pages to go on it uses your site’s own description, and failing that the preset’s opening line. Edit either box, or leave them: Looks right saves both, and Keep the preset (Skip on a twin, which has no preset) settles the question without writing anything. The preview beside the form shows the greeting as the bubble your agent will open with, live as you type.

This question is asked here and not earlier because it could not honestly be asked earlier: until the pages are read there is nothing to write it from. It is also best-effort — Go to your agent stays live the whole time, nothing waits on it, and if the draft is slow the card falls back to the preset’s own words rather than holding you up. Leave without settling it, or reload the page, and your agent keeps whatever the preset gave it, which for a twin is no opening line at all — you can write both any time from its Profile tab.

Training. Reading is not indexing: pages that have been read are stored and counted, and cannot be answered from yet — training is what changes that. Your agent’s first train starts by itself when the read finishes. It runs on our servers, so it carries on if you close the tab, and it is free: a first train never counts against your month. The step row says Training · Under way, with no button to press. Your agent answers from your pages as soon as they are indexed — the line says Your agent answers from your pages now — while it writes its summary of them. Once the introduction is settled, Training your agent takes the screen with its stages, each lit as the job reports it — never on a timer. When it finishes the header reads Trained, a receipt says what the agent now answers from, and Go to your agent is the button on that receipt. The side panel says the same: your agent answers from them. See A new agent trains itself.

If the import could not start that train — rare: the agent already had an index, or the queue refused — the row says Ready to train · Waiting for you instead, and the card beneath the stream asks Train on what you imported? with the figures — sources waiting, words to train on, trains left this month — and one press; how a train works sits behind a How trains work link. Choose what to train opens the fetched sources as tick rows in the same card, each with its size — untick anything it should not answer from and the figures re-count in front of you; the ticks ride with Train now. The card mentions tonight’s automatic train only when overnight training is switched on, and a new agent starts with it off.

You can leave at any point. Go to your agent is on the screen in every state, withheld only for the seconds the create request itself is in flight. It now lands you on that agent’s Train page rather than its Overview, because training is the next thing that has to happen and reading has already finished by the time you press it — and arriving there starts nothing. The train the import started carries on by itself; if none was started, you confirm it there, with the count in front of you. Overview is one click away on the rail and carries the same report of this import.

The ✕ is the other way off this screen, and by now it neither asks nor throws anything away: your agent exists, so there is nothing left to abandon. It goes where it goes from every other screen — your dashboard, with the new agent selected — so it is the exit to take if you would rather look at your account than at the agent you just made.

Either way the reading is a job on our servers: it finishes whether this tab is open or not, and your agent’s own Overview shows the same three steps, from the same place, for a day afterwards. Closing the tab loses nothing.

There is no Back here. The press that made your agent has been spent, and a button that walked you onto it again would be a button that cannot fire.

What a Soon chip means#

Every chip marked Soon is exactly that: noted, not wired, and never a button that quietly does nothing when you click it — no Soon chip is clickable at all. You’ll meet them on three screens: Notion, when you give an address; the tools you don’t use yet, folded under “N more on the way” when you name what you use — every one of them except Cal.com and Calendly on a public twin, where a booking link works today, and every one including those two, under “All N on the way”, on a Support Agent or a private twin; and WhatsApp, Slack, Email and Phone, when you choose where it lives. Dated 2026-09-04, so a later reader can tell when a “Soon” turned into a working connector.

What a preset sets#

Choosing a Support Agent preset on screen 1 is a starting point, not a mode. It writes four things and then gets out of the way:

• A few record-list categories to fill in later — the same ones you could have picked by hand, with the same fields.
• What it says it covers, and how firmly it declines everything else — the scope line you can read on the card before you choose it.
• An opening line for the agent, editable on the first-impression screen before any visitor sees it.
• The sentence it says while handing over, which comes after the admission that it doesn’t have the answer, never instead of it.

Two things a preset deliberately doesn’t do. It never seeds suggested questions — a suggested question is a claim that your agent can answer something, and that claim is checked against your sources rather than assumed. And it isn’t stored anywhere: no screen ever says “this is a Careers agent”, because after creation everything the preset did is visible and editable under Rules and Lists like anything you set yourself. See Four presets, one agent.

How Settings is arranged#

Everything you can change afterwards is under Settings, behind four pills — Profile, Rules, Follow-ups and Notifications. Rules is a list of groups, folded until you open one; the row above each says what it is currently set to, so you can read the page without opening anything. A link from the Overview or a guide opens the group it means. Each group’s How this works opens the matching part of this guide in a panel on the same page — nothing you have typed is lost — and a middle-click still opens it in a new tab.

The personality interview is the first of those groups, a group like the rest with a row that folds it. It used to sit bare above the whole list, which meant a page of settings opened on a chat nobody had come for.

Which groups you see depends on what you made. A Support Agent has no personality interview — it answers in a neutral product voice — and neither a Support Agent nor a Team Agent has Follow-up requests, because neither may ask a visitor for their details at all: a switch that could be turned on and could never do anything is worse than no switch. A Team Agent keeps the interview, because it still answers in your voice.

One Save changes covers the whole page, and it sits on a strip that appears as soon as anything is unsaved and names the groups holding an edit. The five answers below are the exception — each of those saves on its own.

Answers you can still give#

Three questions the create flow no longer asks live on, editable any time, under Settings → Rules → How hiy is set up: its main job, who talks to it and roughly how many, and whether there’s a team. None is required — each has a neutral default that’s simply what happens until you answer it: nothing is promoted and your dashboard keeps the order it ships with; your twin speaks plainly, in English, to whoever arrives; and it’s just you, with team surfaces staying hidden until you say otherwise. The same settings group also lets you record where your material lives — changing that there notes the fact but does not start an import; add material from Knowledge for that.

A change here takes effect the next time the page loads, and it is not a retrain: these answers shape how your twin is set up around you, not what it answers from. Three other answers are not editable afterwards — what you’re building, the tools you use and where it lives are set once, in the create flow, because each of them decided a cap, a chrome or a model tier at the moment the agent was made.

The address cannot be renamed#

Your address on hiy is fixed by the press on screen 5, and there is no rename afterwards. Until that press it is only an answer like any other: screen 2 is where you set it, and stepping back onto screen 2 lets you change it as often as you like. The preview beside the form shows it without offering to change it. Once the agent exists, nothing changes it — renaming isn’t a control this product offers today, so screen 2 is where to check it.

What this is not#

It is not a progress bar. The screens that ask are over in about a minute; the thing that tracks how far along your twin is — claim, source, test, teach, publish, share — is the setup bar and the Set it loose panel on that agent’s own Overview, and both work from what has actually happened rather than from what you’ve answered. See Getting started.

Making a second twin#

A second twin runs the same six screens as the first — including the Back to your dashboard button on screen 1, since now there’s somewhere to return to. Whether there’s a team is answered once for your whole account, under Settings → Rules — none of the screens ask it. Reloading partway through never starts a second agent: before the press there is nothing to duplicate, and after it the address bar already names the one you just made.