What we're building.

Guided twin creation

Sources — paste, blog, YouTube, files, LinkedIn/CV, bulk import

An assistant's memory

Paste what ChatGPT, Claude or Gemini already knows about you, review it fact by fact, and it feeds your twin's answers — without shaping how your twin sounds. A one-time paste, not a sync.

Persona interview that captures your voice

A support-first homepage

The front door leads with the Support Agent: three value cards directly under the hero — the answer with its source, the honest gap, the question reaching the Inbox — each a miniature built from the product's own components, with the hero's strands drawn behind the headline. The Expert Twin remains a clear secondary path.

Launch kit at the moment you publish

The link, a QR code and ready-to-post announcements — three once your twin has a suggested question to lead with, two before that. While an identity check is running you get the link to keep and nothing else: no QR, no posts, no share row. A support or team agent gets the same moment with what it can actually use: who can reach it, the widget snippet or the invite flow.

Starter questions, drafted from your sources

A new twin's page opens with questions it can actually answer, drafted from your own material and yours to edit. Questions you've written are never changed.

Cited answers with honest “I don’t know” gaps → teach → reindex

And when it doesn’t know, it shows the searches it ran first — your question, then its own rewordings. The count never hides, on any plan.

When a rebuild fails, Knowledge says so

A rebuild that runs out of retries no longer disappears into a log. Your Knowledge page says how old the index your twin is still answering from is — so you know exactly what is missing — and offers a retry. Your twin never goes offline for it.

Verifiable receipts

A citation opens to the whole passage, with a link that jumps to it on your original page. hiy writes its own summary of your material to answer well — that summary is never quoted to a visitor, and no longer counts as one of your sources.

Source controls

Show the sources behind answers or keep them private; optionally surface the exact passage up front as verifiable receipts. The AI label, the honest “I don’t know” and its search count, and the report link never hide.

The hiy.ai credit

The “Powered by hiy.ai” line comes off on Founding, and you can keep it if you’d rather be credited — while it’s there, anyone who follows it and signs up is recorded as your referral. The AI label, the honest “I don’t know”, citations, the report link and the Verified badge can’t be removed by any plan.

Light and dark mode

A toggle at the top of marketing and docs pages, and in your account menu once you're signed in. The choice applies across the whole site and is remembered on that device — hiy still opens dark by default rather than reading your system setting, because the grainy wallpaper is too much of the point to leave to it.

The Sandbox

A private sandbox to try your twin before publishing — nothing counted.

Saved questions

Keep the questions that matter and say whether each should be answered from your sources or honestly declined. The set re-runs itself after every train, on the server, whether or not the page is open — a finished train is already answering visitors, so the check has to happen without anyone remembering it. Each run is a real answer from your real index, and the result is a count of what changed — never a score, and never a pass rate.

Scope boundaries

It answers on your subjects and declines the rest, instead of acting like a general chatbot.

Follow-ups & People

Visitors can leave their details; you get a People list with context and CSV export. New twins ask for an email after a couple of messages — one tap turns it off.

Lead pipeline

Say where each follow-up stands — New, Replied, Booked, Won or Closed — filter by any of them, and see how fast your replies go out. Your own marks: hiy sees a booking-link click and will never promote it to a booking on your behalf.

Lead & booking attribution

Open the conversation behind any lead or booking-link click, with the moment they left their details marked inside it. Team Agents are deliberately excluded — no per-person trail.

Email notifications

A daily or weekly summary that leads with follow-ups and booking-link clicks, plus optional instant alerts. All optional.

Lists your twin looks up

Keep services, open roles, products or common questions as a list, with your own fields. Your twin searches them and answers from the exact rows rather than from memory — and books with your Cal.com or Calendly link.

Public page, inline embed, share, export & delete

Channels workspace

Configure the Chat widget and the Embedded ask bar from one global place, with a direct link to the existing hosted page for personal twins. Reuse a workspace template, choose a different one for an agent, preview draft changes live, and publish only when ready.

Team Agents

You, answering invited colleagues from internal notes — reporting what you wrote, never deciding for you. Invites are just email addresses.

Support Agents

An agent for your product inside your own site: neutral voice, token-gated, and it hands what it can't answer to your inbox.

Invited colleagues can read the Inbox

A support agent's handoffs, and the conversations behind them, open for anyone invited into the organisation — not only the account that created it. Reading only: they can't change the agent, publish it or touch billing. Team Agents are excluded, because an inbox of who-asked-what is the per-person trail they promise never to keep.

A customer-support product line

Support Agents have a product face: a Product menu across this site, a page each for the Support Agent and the Team Agent, and a trust page covering identity review, token security and who processes what. At capacity — the monthly cap or a rate limit, deliberately indistinguishable — an agent says so in the server's own words and still offers the handoff form, keeping the question the visitor typed; a handoff emails them a receipt, which states a typical reply time only when there is a real one to state; the review queue and any alert email that fails to send are watched rather than assumed. Visitors can mark an answer helpful or not. Creating one offers four named presets — SaaS product support, Orders & billing, Docs/developer assistant, Careers/HR — each a starting setup (record-list categories plus a declared scope) over the same single Support Agent, and all of it editable afterwards; a preset is not a different kind of agent.

Removing a person erases their conversation

Taking someone out of People now takes their transcript with them, not only their email address. You can also delete a single conversation on its own, owner-only, from its review screen in Activity.

Embeds: inline, and a floating bubble

One snippet either way. The bubble floats bottom-right and loads nothing until a visitor opens it.

Social cards for shared links

A pasted link unfurls with your name, your role, and what your twin promises — generated from your profile, nothing to upload.

Conversations survive a page refresh

A visitor who reloads keeps the thread they were in — and closing the tab ends it. The transcript lives in the tab, not on the machine and not behind a session id anyone could ask hiy to replay, so a twin opened on a borrowed laptop leaves nothing for whoever sits down next.

One library, shared between your agents

Your material belongs to the account, not to one agent. Knowledge Hub lists every source you hold, grouped by the site and section it came from, with which agent it was added on and which agents draw on it. It has an Add of its own, and what you add there belongs to the library rather than to any one agent — each agent picks up what it needs by searching the hub from its own Add, and anything you add on an agent appears in the hub tagged with it. The hub has an allowance of its own, five times what one agent holds. Nothing is copied and using a source on a second agent costs nothing, so fixing your bio once fixes it everywhere. A source can be marked confidential, which blocks it from ever reaching a public twin — enforced in the database, in all three directions it could be reached, and on the read as well as the write.

Delete one agent, and only that one

Delete one agent without touching the others. The control names it — on the agent's own Profile and as a per-agent list in Settings — and before you confirm it says exactly what goes: its conversations, its leads, its answer index, and its sources by name, split into the ones that die with it and the ones another agent keeps, which move into your Knowledge Hub instead. It asks you to type the agent's own name, not the word “delete”. And your address is held for you for 90 days afterwards, so a link or embed pointing at hiy.ai/<name> can't be taken over by a stranger and you can use it again yourself — deleting your whole account holds its addresses the same way, and Settings → Data & privacy lists what you're holding with a Release beside each one. Immediate and final: no grace period. Two things outlive it on purpose and the screen says both — material another of your agents uses moves to the Knowledge Hub rather than being destroyed, and the address it answered at is held for 90 days so a saved link cannot be taken by a stranger.

Sync a site you've already imported

Web pages change after you import them. Sync, on any site in Knowledge Hub, re-reads every page you already hold from it and replaces the text where the page changed — a title you renamed yourself stays yours — then asks the site what else it has and offers anything new in a checklist. Nothing new is added unless you tick it. You get a report either way: what changed, what is new, and any page that couldn't be read, with the reason. A page that fails is left exactly as it was; a sync never deletes anything. And it doesn't train — the new material waits for the train you confirm, or for the overnight one if you have switched it on.

Four hundred sources you can actually read

Import a site and your sources stop being a list you can scan. So they aren't one: they're grouped by where they came from — the site, then the section of it — biggest first, each group saying how many sources and how many words it holds, so the thing you have too much of is the first row you read. Search matches the title, the type and the address, and opens every group it keeps. Chips filter to one kind — web pages, video, files, notes. A short list keeps the plain list, because there is nothing there to find. Words used is now stated once at the top of the screen instead of beside every tab, and what an agent refuses, covers and hands over is its own Guardrails tab.

More than one twin

A public twin for the world, a Team Agent on internal material, Support Agents for your products — each with its own knowledge, audience and rules. Your public twin is yours alone — and what stops anyone publishing a twin of someone else is the identity check, human review, and a report link no plan can remove.

Private mode

Unlisted links — reachable by link, out of search, sitemaps and listings.

Identity guardrails

You may only twin yourself. An automated check clears your own name in your own material instantly; anything else waits for a person who sees why it was flagged. Saying no is always a human decision.

Verified badge

That identity check, made visible: a live twin carries a Verified badge on its page, so a visitor can see it represents its real creator before trusting a word it says.

Insights: activity, top questions, and the gap queue

Coverage sits on your Overview rather than here — this week's questions across every agent you own, minus the ones it said it couldn't answer, and never a claim about whether an answer was right, or where it came from — and its shortfall is a link straight into the gap queue on Insights.

Deflection and escalation, shown

Insights opens with the two rates a support agent is judged on, counted from what the server itself observed rather than anything a browser claimed — with the conversations behind each percentage printed beside it. Below twenty conversations it states no percentage at all: “Not enough conversations yet”, and how far along the sample is, which is the state most twins are in. A window too busy to total exactly says that instead of approximating. Neither rate claims a visitor's problem was solved — hiy does not observe that and will not model it — and deflection is never a claim that nobody was handed off.

Escalation you set

A Support or Team Agent's handover is yours to decide, on an Escalation page: when it hands over — when it can't answer, that plus whenever somebody asks for a person, or never — who the handover reaches, and the sentence the visitor reads while it happens. One policy per agent, deliberately not an ordered rule list: a list is where a workflow engine starts, and the promise was a policy you can read in one breath. Saving is prompt-layer, live on the next question, with no rebuild and no training spent — editing a sentence must never cost part of a monthly allowance. hiy does not know your working hours and will not guess at them, so a handover is always attempted and nobody is told the team is offline. A personal twin has no such page: it answers as one person, so there is nobody behind it to hand a conversation to.

Upgrade prompts that wait for something to have happened

It only ever asks while there is a plan to buy — and then hiy mentions Founding once, after something has actually happened: a follow-up, a booking-link click, a month's messages used. Never more than once a fortnight, never on a timer, and never inside the conversation it affected: a twin that has run out of messages says so to the visitor, and the card about it waits on your Overview. Dismiss three and it stops for good; the switch that turns it off entirely is in Settings → Plan & billing today.

Per-plan usage metering

Words, sources, monthly messages and monthly trains, each metered against your plan and shown to you. The message allowance is one pool per account, shared by every twin in it — enforced and displayed as the same number, so the meter and the wall can never disagree.

One wallpaper-glass design system across the site

The grainy gradient and its translucent panels carry every page, with two deliberate exceptions: the embed, which renders inside other people's sites and stays neutral, and the docs, which run a flat shell of their own because documentation should read as documentation.

Free tools

A repeat-questions calculator and an audience-questions generator — free, no account needed.

A dashboard that arranges itself

The sidebar is grouped into the five stages of getting a twin working — Build, Train, Test, Publish, Measure — and that order is the same for everyone on every visit. What follows what you actually open, plus the five setup answers below, is the order inside a stage. Nothing ever disappears, nothing leaves its stage, and the top of each section never moves. It is a table you could read, not a model scoring you.

Your Dashboard and your agents' Overviews are two different pages

The first screen is about your account: every agent as a card with a setup bar showing how far through the six activation steps it is, the week's numbers, your plan and how many days of your trial are left, one queue of everything that needs you, and four quick links. Open an agent and its own Overview carries what belongs to it — how complete it is, what to do next, whether anyone can reach it, and the box you answer its unanswered questions in. That box in particular had to move: answering writes a correction to one agent, so it belongs on the page that names it. Nothing on either page reorders itself any more.

A console for the app

The dashboard and the guided start are now a console rather than a page of panels: a narrow spine of three sections — Dashboard, Agents, Knowledge — that opens to words the moment you point at it or Tab into it, and pins open if you would rather it stayed; a second rail and a docked preview chat that appear for Agents; a top bar carrying the agent's name, the preview toggle and Publish; and one page header shape everywhere. Activity is every conversation as a run: what was asked, which agent took it, and how it ended — answered, a lead, handed over, or declined — over 24 hours, 7 days or 30, with a review screen behind each one holding the transcript and what the answer used. Build is one hub over the six surfaces an agent is assembled from. The app opens on a light skin; the marketing site and the public pages are untouched, and a stored dark choice still wins. Where a number is not stored, the console says Soon in its place rather than inventing one.

Knowledge you can see the state of

A source appears the moment you add it, and because changes wait for a train — Train now, or the overnight one if you have switched it on — the row carries a Waiting chip until one lands, so an “I don’t know” straight after adding something is explained rather than mysterious. A source that is live is marked with nothing at all: the state worth interrupting you for is the one where your twin cannot answer yet. An agent’s knowledge is now four panels you can link to on their own — Sources, Add, What it knows and Guardrails — with one source list instead of two, and the account-level Knowledge Hub has two of its own, Sources and Add. And a twin with nothing in it opens on four starting points that take the paste, the link or the file on the spot, instead of pointing you at a tab.

Training you can see, and that happens without you

Training is a stage of its own. Changes wait for a train rather than rebuilding the moment you add something, so nothing your visitors get changes without you, and the sidebar carries the number of changes waiting. Train now starts one, and overnight training is a choice: a Train overnight switch, off unless you turn it on, hands that job to the night in your own timezone — with it off, what is waiting waits for you. Before one starts you get a count rather than a guess: what is being added, what your twin will hold afterwards, and the line that a train rebuilds everything — ticking chooses what your index contains, not how much work runs. Untick a source and it is held back until you tick it on again, out of every train and costing nothing against what your plan holds. A finished train is live: checking happens after it goes live, not before it. And when a month of trains is used up, learning pauses and says so with the date it resumes — while a train that takes your own material down is never paused.

A guided start

Six screens, one typed field: what you are creating — a twin, public or private, or a Support Agent — then one address to read from. The next screen shows everything found under that address before a single page is fetched: grouped by path with counts, a filter, a running total, and a one-press untick for the groups that look like duplicates — a language version, an archive listing, a numbered list page — each saying what it actually matched rather than claiming an AI suggested it. Pick twenty pages instead of importing four hundred, and the import runs while you answer the rest. From there it fills in a name, a greeting and what it covers — with the photo and accent already taken from your own site's icon and theme colour where it declares them, and yours to change — asks which tools you use and where it will live (each with a plain Soon for what is not wired yet), and finishes on that agent's own Overview, where the import reports on itself. Back is on every screen but the first and keeps what you typed, in that tab; it never un-makes the agent or re-runs the import. The flow keeps no progress of its own: the Set it loose panel is the one thing that tracks how far along your twin is, and since 2026-09-04 it draws that as six weighted segments and a percentage — each segment as wide as the stage is worth, so what you see filled is the number — going on what has actually happened rather than on which screens you walked past.

One state word, from one reader

The Sandbox, the top bar’s Publish control and the agent’s own pill in the rail take a state from the same reader instead of each naming it their own way, so two places on one screen can no longer disagree about the same agent. The word says what the system knows — Live, Waiting for a train, Nothing to answer from yet — and never how good the answers are.

The create flow opens on your account’s own facts

Before the first field, screen 1 prints what you already hold: the public twin you already own and its address; where it bears on what you are about to make, how many of your three agent slots are used and how many of those are unfinished drafts holding one; and whether the Founding trial is already spent. A refusal you can see coming is not a refusal.

Voice & rules says what is still only on screen

An Unsaved changes bar names the groups holding an edit and carries the Save itself, so the button sits beside the thing it saves and a refused save says why on the bar that refused it. Discard asks nothing first, because every edit it drops is still in front of you.

Insights and Activity say which agent they are about

On Insights every block that reads one agent’s figures says so in a line of its own, with an agent picker beside that line once you hold more than one agent; Activity names the agent on each run whenever the window covers more than one. Both keep a URL that matches what rendered rather than the view you arrived from, so a link you send opens on the page you were looking at.

The knowledge library says when each page was last read

Every row in the account’s Sources list carries the date of its last read, and the rows that came from a web page carry a Re-read that fetches just that page; a row nothing has ever read says so rather than showing a blank. The stamp moves on a re-read and not only on an add, so it means last read and not first read. The agent’s own Sources rail does not print it yet.

A handover can be marked replied

An invited colleague who can see the conversation can record that they answered it, and the row then says who and when, so two people don’t both pick it up. It is one status write and nothing more: handovers still can’t be assigned to a named colleague.

Documentation that reads like documentation

The whole help set at /docs: a sidebar grouped by what you are trying to do, an “On this page” contents rail beside every article, code blocks highlighted at build time with a copy button, and search in the bar above every page — ⌘K, or / — that lands on the heading which answers you rather than the top of a long page. The index starts with four paths for the jobs readers actually arrive with, then lays out every guide in the same numbered sections as the sidebar, with 44px navigation rows that remain comfortable to tap on a phone.

Actions

An agent can collect specific details in the conversation, and post a short summary to a URL you choose. Credentials are stored encrypted and every use is recorded. Previews rehearse writes rather than making them, and the answer says so. Two actions per answer, one call each.

Facts your site can tell the agent

The token your own server already mints for the embed may carry up to six facts about the person asking — an order number, a plan name — under key names you declare on the agent's Publish tab beside the signing secret — on an agent that takes them, which today means a Support Agent. Your site cannot invent a field, and declaring one never obliges your site to send it. It makes “where is my order” actionable, not conversational: the values reach the action and nothing else. The agent cannot weave one into a sentence, cannot greet somebody by an attested name, and cannot mention one in an answer that ran no action — it is told only which facts arrived, by name, and never what they say. The visitor is told too: a short line under the answer names the keys that arrived, never the values, and no plan or setting removes it. Nothing is written down — values are read for the length of one answer and reach no conversation, no follow-up record, no event and no index, so there is nothing for a deletion to reach. What your own endpoint sends back is different: that is your material, kept with the conversation like any other answer.

Mind — a map of what your twin knows

A 3D scene of your twin's material, one point per passage, coloured by the source it came from and placed by how alike the passages actually are rather than sorted into folders, with lines joining the ones your twin reads as being about the same thing. Orbit it, zoom in, and click a point to inspect its source; an accessible List view and an owner-only View by: Sources | Groups filter lens are available too, with Groups reusing the Sources panel's origin groups. It's built from your twin's live index: what's on the map is what your twin can actually answer from, and a new source joins it at the same train that makes it answerable, never before. Sharing it is opt-in and off by default — turn it on and a visitor to your public page can orbit the same shape, positions and source types only, never a title, an address or a word of content, and a source you've held back or marked confidential never appears, shared or not. It's on your public twin first; Support and Team Agents don't have it yet.