Trust, stated as rows you can check.

Every row links to the doc or policy behind it, or says it is enforced in code.

01 · Identity & review

Identity, and who reviews it.

The identity policy is enforced by machinery, not stated in a FAQ.

Where an agent can be

StateA stranger getsMoved there by
DraftNot foundWhere every agent starts
Awaiting reviewNot foundYou, by publishing
LiveThe agentThe screen, or a person
DeclinedNot foundOnly a person, with a reason
Create · Address

I confirm this agent represents me, or a product or organisation I own — and that I have the rights to the content I’ll add.

Unticked, the API refuses“You must confirm this agent represents you, or a product or organisation you own.”

Overview

This agent wasn’t approved

A reviewer couldn’t confirm this agent represents you or something you own.

You may only twin yourself, or a product you own.

One rule in the terms, the API and the checkbox. No “with permission” exception, because permission cannot be verified.

The identity screen can approve; it can never reject.

A deterministic screen, no model call. It approves only a name matching the account’s own identity with sources about that person, or a Support or Team Agent in an invited Early Access workspace. Everything else waits for a human, and rejecting stays a human act, with a written reason.

By design, enforced in code

A decision is made about the exact name the reviewer read.

Approval records the name the reviewer saw and refuses to land if the agent was renamed meanwhile, so a cleared name never becomes a different person’s.

By design, enforced in code

The Verified badge is derived, never stored.

It follows the cleared name: rename an approved twin away and the badge goes; rename back and it returns. A rename launders nothing.

On failure:

When the screen cannot clear an agent, the agent waits.

No timeout publishes by default. The agent stays at awaiting review, a 404 to strangers, until a human decides: publicState in lib/agent-access.ts agrees with REVIEW_ENFORCED. The queue is admin-only, not a public status page.

By design, enforced in code

The states are drawn in what happens when you publish.

02 · What no plan removes

What no plan can remove.

These surfaces are never removable on any plan; only the “Powered by hiy.ai” credit can go.

yoursite.comExample
Yoursite support

Can you move me to yearly billing?

Yes. Yearly billing starts at your next renewal, and I’ve asked the billing team to switch your plan.

Billing FAQdocs.yoursite.com/billing
  • Billing change requestSent to the team.
From the site you're on:plan

Do you take bank transfers?

I don’t have the details on bank transfers here, so I won’t guess.

Searched 3 ways· See searches

AI agent · may make mistakes

Report · Processed by hiy.ai · Privacy

On every plan, on every answer

  1. CitationsThe source behind a grounded answer
  2. The action receiptWhich action ran: “Sent to the team.”
  3. What the site told this answer“From the site you're on:” and the keys, never the values
  4. The honest gapA dashed card that says so, and how many ways it searched
  5. The AI label“AI agent · may make mistakes”
  6. The report linkBeside who processes the conversation
1on the picture.

Citations behind grounded answers.

No plan can remove them. The owner chooses how they render, or hides them from visitors; paying cannot make an answer look better-sourced than it is.

2on the picture.

The action receipt.

An answer that ran a Notify or a Call your endpoint names the action and how it went. No plan gates it, and the setting that hides sources does not reach it.

3on the picture.

What the site told this answer.

When your signed token attests facts about the person asking, the answer names the keys that arrived, never the values. A visitor is told what an answer used about them. No plan or setting removes it.

4on the picture.

The honest gap, and its receipt.

It says so instead of inventing, and shows how many searches it ran. The count shows on every gap, on any plan or setting.

5on the picture.

The AI label.

Every agent is labelled AI wherever it answers, and says so when asked. De-branding swaps hiy’s name for “AI twin”: more disclosure, not less.

6on the picture.

The report link.

Every public twin carries one. Reports go to an operator queue, not an unread mailbox.

By design, enforced in code

Cannot:

A hidden citation cannot be recovered from the page.

When an owner hides sources, citationsForVisitor in lib/rag/tools.ts removes the citations, and the source id, from the response before it is sent. “Hidden” means absent, not just not shown.

By design, enforced in code

03 · Who processes what

Who processes what. And for how long.

Who answers for which data, who touches it, and for how long.

Who answers for which data

  • Your agents' material, conversations and contacts

    You · controllerhiy · processor

  • Your account: email, billing, product events

    hiy · controller

  • A fact your server attests in a token

    You · controllerhiy · keeps nothing

The services that process it

Vercel
Hosting, AI gateway
Supabase
Database
Anthropic
Models
OpenAI
Models, search
Google
Models
Fireworks
DeepSeek models
Baseten
DeepSeek models
Microsoft Azure
DeepSeek models
DeepInfra
DeepSeek models
DigitalOcean
DeepSeek models
Cloudflare
Network, page reading, bot check
Resend
Email
ElevenLabs
Voice
Telnyx
Phone calls
Meta
WhatsApp
Stripe
Payments
Settings · Your data

Your data

Your content is yours — take it or remove it whenever you like.

Export my data (JSON)
{
  "agents": [ … ],
  "sources": [ … ],
  "conversations": [ … ],
  "messages": [ … ],
  "held_handles": [ … ],
  "org_credentials": [ … ],
  "credential_audit": [ … ]
}

Processor for what your agents hold; controller for your account.

What you put in and what your agents collect, conversations and contacts included, is yours: you control, hiy processes, under the Data Processing Agreement. Your account, billing, product events and security records: hiy controls. A fact your site attests about a user is yours, and hiy never keeps it.

The services that process your data, named.

Vercel serves the application and, through its AI Gateway, reaches some models only on zero-data-retention hosts, from an allowlist: Google's own for Google's models, and Fireworks, Baseten, Microsoft Azure, DeepInfra or DigitalOcean for DeepSeek's. Supabase hosts the database and authentication in the EU (eu-west-1). Anthropic and OpenAI run their models directly; OpenAI also makes the search embeddings. Cloudflare carries every request to hiy.ai on its network, opens pages that need JavaScript (getting only the address), and runs the bot check on sign-in, public forms and free tools, which gets the visitor's network address. Resend sends notification email. ElevenLabs carries voice calls for workspaces in the voice pilot, on the website and by phone, and may process a call outside the EU. Telnyx carries calls to a business's hiy phone number, in the voice pilot, and may also process a call outside the EU. Meta delivers WhatsApp messages, in test, when an owner connects a WhatsApp Business number. Stripe processes payments, if you subscribe. Each processes data only to provide its service to us.

Kept while your account exists, then really gone.

Delete an agent, a person or your account and its data goes with it: deletion, not a hidden flag. Three exceptions are named. Deleting one agent moves material another of your agents draws on to your Knowledge Hub instead of destroying it. Agents you made in a workspace someone else owns stay there, with its owner, when you delete your account. And each deleted agent’s address is held for 90 days, the name and the date only, so a saved link cannot be claimed the moment you leave. A fact your site attested is outside this promise: it is never written down, so deletion has nothing to reach.

Action credentials are stored encrypted, and every use by an action is recorded.

An action that writes into your help desk needs your key, so, unlike a signing secret, it is kept: in Supabase Vault, where the database holds ciphertext and the key that opens it is read from the server at startup, so a dump, backup or replication stream is unreadable to hiy on its own. Inside the application, every read goes through one function, read_org_credential in supabase/migrations/20260831090000_org_credentials.sql, which checks the credential is yours and unrevoked and logs which action asked and when; that log is in your account export. Replacing or removing a credential destroys the old ciphertext. What that does not protect you from: hiy’s own service role can call that function, and anything holding the service-role key and a direct SQL connection reads the values straight out of the vault without passing it, so nothing is written to that log. It is a record of what your actions did, not proof of everything that could have happened. Encrypting at rest defends the database, not the application. Give an agent a key scoped to its action, and remove it when done.

The privacy pointer survives de-branding.

Every conversation carries a “processed by hiy.ai · privacy” link in its chat chrome, as never removable as the AI label. Paying removes hiy’s marketing credit, never that pointer.

Cannot:

A Team Agent cannot tell an admin who asked.

api/twin/conversation answers 404 for any agent whose variant is internal, so no one, owner included, can open a team agent’s transcript, and the People view drops attribution for those agents. It cannot ask either: allowCapture: false on the team-agent policy caps the owner’s follow-up switch, and mayCapture in lib/agent-kinds.ts composes the two wherever a card could appear. It cannot be told: allowAttested: false on the same policy means signed facts about the colleague asking are never built for it to read. An admin sees topics and gaps, not who raised them.

By design, enforced in code

Cannot:

hiy cannot join your two visits.

Your token carries your own site’s id for the person asking. hiy compares it once, to recognise an owner previewing their own agent, and never writes it: the conversation row has no column for it, and the column and its index were dropped. “Show me everything this person ever asked” has no answer to give, on any plan and for any operator.

By design, enforced in code

A Data Processing Agreement on the EU's standard clauses.

The DPA is the European Commission's standard contractual clauses for controllers and processors, word for word, with the choices made and the annexes filled in for hiy. It forms part of the Terms.

Technical detail

04 · Token security

Token security. No token, no agent.

No public URL: visitors reach it with a short-lived token minted by your server.

A visitor token

Signed with
HMAC-SHA256, and nothing else
Lasts
Short-lived: you set it, one hour at most
Refused beyond
60 minutes, however well it is signed
Opens
One agent, never another in the same organisation
Signing secret
Derived per organisation, stored nowhere
May carry
Up to 6 facts the owner declared, written nowhere
Without one
A 404, the same as a made-up address

Signed with HMAC-SHA256 — deliberately not JWT.

One algorithm, no header to confuse, no “alg: none”, no key discovery. A few-minute credential for one agent needs no federated-identity framework.

Lifetime is capped, and the cap is ours.

Tokens default to fifteen minutes and are refused beyond MAX_TOKEN_LIFETIME_SECONDS = 3600 s (one hour), even when signed correctly.

By design, enforced in code

Audience-bound.

A token minted for one agent never opens another, even in the same organisation.

By design, enforced in code

Signing secrets are derived, never stored.

Each organisation’s secret derives one-way from a single master, so none is written down and one reveals nothing about another. Credentials for actions are stored, encrypted: see below.

By design, enforced in code

A token may attest a few declared facts, and hiy writes none of them down.

Up to six facts about the person asking, only keys the owner declared. readAttested in lib/attested.ts reads them for the length of one answer and writes them to no row: not the conversation, follow-up record, event or search index. A value your endpoint sends back is read only from a field the action declares, and an answer using it is kept with the conversation.

Verification fails closed, and explains nothing.

The signature is checked in constant time before parsing, and a caller holding a guess is never told which check failed. Only a token genuinely minted for that agent hears “expired” or “wrong site”.

By design, enforced in code

Without a token, the agent does not exist.

Asking returns the same 404 a made-up address does, not a login wall that signals something to attack; Team Agents answer outsiders the same way. A Support Agent opened on its owner’s sites gets a hiy-minted token for requests from a listed Referer, a check browsers honour and programs do not, so its limits are on cost, not on who can ask. A Team Agent can never be opened; the database refuses it.

A token hiy mints carries no facts, and can only look things up.

hiy’s own minter attests nothing, so an open visitor is anonymous. The only action offered is a look-up needing no stored key: none that uses your key or site facts, sends data or posts a notification, decided per answer.

Drawn in how a token is built and checked.

05 · Limits that fail closed

Limits that fail closed.

When metering or entitlement fails, it fails toward not serving, never unmetered.

Where each failure lands

Ifhiy
Two conversations race for the last messageOne gets it. The cap is checked and counted in one step.
The message counter errorsThe answer is declined, never served unmetered.
Reading the plan failsThe Free plan’s limits apply, not the paid ones.
The nightly billing job never runsThe plan is still read from the trial’s dates, not the label the job writes.
The month’s messages run outThe visitor is told so, never shown a dead widget.

The message cap is checked and counted in one atomic step.

Concurrent conversations cannot race past it, and if the counter errors, the answer is declined, not served unmetered.

By design, enforced in code

Entitlement is re-derived on every read.

The stored plan is only a cache. A trial’s end is read from its dates even if the expiry job never ran, and a failed read serves the free plan.

By design, enforced in code

Hitting the cap is stated, never disguised.

Out of monthly credits, an agent tells visitors and stops; an embedded Support Agent at its cap states the limit and offers the contact form. Never a dead widget, never an auth-shaped error.

On failure:

When the billing cron fails, the plan still follows the dates.

The nightly job only tidies the plan label. lib/billing/entitlement.ts derives the answer from the reason columns: a trial keeps its plan until its end date (or, if it ended before paid plans opened, a short grace after they open), and a Stripe status outside the entitled set loses it, job or no job. A missed run gives nobody a plan they should not have.

By design, enforced in code

Availability, honestly.

A public status page shows whether hiy is working right now. No availability percentages: we won’t publish figures we aren’t measuring against a public record. The service is provided as-is.

06 · How it’s engineered

How it’s engineered. Read from the code.

No machinery diagrams, no benchmark charts; any number here is read from the code.

Three prompt layers

  1. Layer oneEvery agent

    Say you’re an AI. Never invent. Admit the uncovered half.

    No setting reaches it

  2. Layer twoEach kind

    A Support Agent never sells; a Sales Agent may. A Team Agent reports, never decides.

    No setting reaches it

  3. Layer threeYours

    Voice, scope, off-limits topics.

    You write it

Which model answers, today

DeepSeek V4 Flash
Expert Twin chat, Team Agent answers, wiki synthesis
Claude Haiku 5.5
Support answers
Claude Haiku 4.5
Suggested follow-ups, gap questions

Routed by task. The default model is the same on every plan.

Read from the code

  • 800output tokens

    The longest a Support Agent’s answer can run

  • 3searches

    The most one answer runs before it answers with what it found

  • 10minutes

    A rebuild’s lease, after which the next run proceeds

Three prompt layers, and only one belongs to the owner.

Layer one is identical for every agent: the honesty rules (say you’re an AI, never invent, admit the uncovered half). Layer two is identical per kind: a Support Agent never sells (only one its owner labels a Sales Agent may); a Team Agent reports and never decides. Layer three is the owner’s: voice, scope, off-limits topics, and the words it uses when it can’t answer. No setting reaches layers one or two.

By design, enforced in code

Retrieval over a synthesized wiki — not a fine-tune.

Sources are rewritten into a structured wiki and indexed. Every answer is grounded in passages retrieved for that question, and a grounded answer can show its passage.

Sources feed retrieval, never training.

Your material becomes an index searched at answer time. It trains no model: not ours, not a shared one, never another owner’s agent.

Models are tiered by task, and the default is the same on every plan.

Where a person’s voice is the product (an Expert Twin’s chat, a Team Agent’s answers, the wiki synthesis behind them), the model is the one that passed our answer evaluation, not the priciest: today DeepSeek V4 Flash, through Vercel AI Gateway on zero-data-retention hosts. Support answers have a default of their own, today Claude Haiku 5.5, because a password-reset answer needs speed and precision, not a voice. Follow-ups and gap questions run a lighter tier, today Claude Haiku 4.5. The promise is the routing, not the vendor, and the default is the same on every plan.

By design, enforced in code

Cost discipline is fail-closed too.

A Support Agent’s answer is capped at 800 output tokens by its entry in lib/agent-kinds.ts, and an answer that searches gets a hard ceiling of MAX_SEARCHES_PER_ANSWER = 3 rounds, then answers with what it found.

By design, enforced in code

Cannot:

hiy’s own prose cannot be quoted to a visitor as a source.

Every indexed passage stores where it came from: buildIndexChunks in lib/rag/chunk.ts stamps origin as source (the owner’s material) or summary (hiy’s synthesis), and every summary passage is dropped before citations reach a stranger. The cost is accepted: a wiki-led answer shows fewer citations, sometimes none. Showing more would present our sentences as the owner’s.

By design, enforced in code

Cannot:

A Support Agent answers from your lists and gives your booking link, but never recommends, quotes or takes details to sell; a Sales Agent, labelled as one, can.

Skills it is never offered, not an instruction it is asked to follow. The catalogue match, quote request and offers skills carry forLabel in lib/skills/registry.ts, so a Support Agent holds them only when its owner labels it a Sales Agent, and allowCapture: false in lib/agent-kinds.ts keeps it from asking for contact details to sell; a handover can still ask for an email so your team can reply. It still reads the lists its owner attached, services and prices as written, and gives the booking link there (readsServices).

By design, enforced in code

Cannot:

The search receipt counts queries; it cannot claim results.

The receipt lists the searches the agent chose to run, in order: queries, not a verdict. hiy never says a search “found nothing”; it says what it looked for.

On failure:

When a reindex fails, the lease expires and the next run proceeds.

A rebuild takes a database lease of REINDEX_LEASE_SECONDS = 600 s (ten minutes), not a lock it must remember to release. A process that dies mid-rebuild blocks nothing: the lease runs out and the next attempt takes it, leaving a visibly stale index, never a knowledge base wedged shut.

By design, enforced in code

Drawn in why only your words are quoted.

Check every row yourself. Start free; Growth is free for 14 days.

Start free
  • 14-day free trial
  • No card needed
  • Paste a URL to start

Check it live: ask our own agent if it’s an AI, then something its material doesn’t cover.

Missing a row? Ask us. If the answer is “not yet”, you get “not yet”.