Privacy

Perfect Human holds health information you choose to give it, so that a team of AI agents can give you advice that accounts for your actual situation. We do not sell it, we do not advertise against it, and we do not share it with employers, insurers, or data brokers.

We are not a healthcare provider. Perfect Human is not a substitute for a doctor, and nothing in it is a medical record.

Only what you enter or connect. Specifically:

• Account: email, name, country, timezone, and optionally date of birth and sex, which the lab-result features need to interpret ranges.
• Your questionnaire answers: goals, symptoms, history, lifestyle, and preferences.
• Your conversations with agents and in group rooms.
• Health data you add: blood work you upload, supplements you track, workouts you log, experiments you run, plans you make, documents you upload, and GLP-1 details if you use that feature.
• Apple Health data you import: if you choose to, resting heart rate, heart-rate variability, sleep, and respiratory rate. The export file is read in your own browser; only the summarised daily numbers are sent to us.
• Usage: pages visited and features used, for analytics (see below).

We are a very small team and we run on other people’s infrastructure. Your data is processed by:

• Supabase: the database, login, and file storage. Hosted in Sydney, Australia.
• Vercel: application hosting.
• OpenRouter: the gateway the AI models behind your agents are reached through. Your message and the relevant context pass through it to the model that writes each reply, and to the company hosting that model. Today that is mainly GLM, a model made by Z.ai, run by hosting companies we name in advance (Friendli, Fireworks, Together and a few others), and Google’s Gemini models when you press “Think harder” or when those hosts are down. We use them under terms that do not let them train on what we send. Their servers are mainly in the United States.
• Anthropic: Claude models, kept as a backup. Your message and context go there only if an agent is switched to a Claude model.
• OpenAI: converts text into the numeric form used for memory search, and reads replies aloud. Text is sent; no replies are generated.
• Qdrant: stores that numeric form so agents can recall earlier conversations.
• Deepgram: turns your dictation into text (more below).
• Resend: sends the emails you allow, such as your digest and notices about your account. It receives your email address and the message.
• Stripe: payments, if and when we charge. Card details never touch our servers.
• Google Analytics and Microsoft Clarity: usage analytics.

If you connect your own AI provider key, your conversations run on your key and your account with that provider instead of ours.

• Health screens are hidden from session analytics. Clarity records how people move around the app, but the parts showing your labs, supplements, GLP-1 details, and conversations are masked out before recording.
• Messages that trip our safety check are stored like any other. This changed on 8 September 2026. Before that, a message at the most serious level was discarded: not saved, not indexed, not added to any agent’s memory. It is now kept, because throwing it away meant the app forgot the most significant thing you had told it and your agent could not refer back to it. Alongside it we keep a note that the check fired, when, and which resources you were shown, so we can tell whether the safety system is working. If you want a conversation to leave no trace at all, use a private chat. Tap the ghost in any conversation, including a group chat, or start a private group. Nothing said in one is recorded. A private chat with one agent is deleted when you close it or after 7 days without a message; a private group is deleted when you close it.
• Dictation: your recording is used to write the words, then discarded. You can speak to your agents instead of typing. When you tap the microphone, the recording is sent over an encrypted connection to our server and passed to a transcription service (Deepgram, or OpenAI if that is unavailable) to turn it into text. It is held in memory for that one request and is never written to storage (not by us, and not kept by them for training). You see the text before anything is sent, and can edit or delete it. If your browser can transcribe on its own and our service is unavailable, it falls back to that instead, and the audio never leaves your device. Feedback you speak works the same way: we keep the words you send, never the recording.
• Voice notes you choose to send are kept. Sending a message as a voice note stores that recording, so you and your agents can play it back later. It is kept in storage only your account can read, and deleting the message deletes the recording with it. Your agents read the transcript, not the audio.
• Words we learn from you stay with you. So that names, dishes and supplements come out right, we keep a short list of words you use: some added by you, some learned when you correct a word before sending. It is stored with your account, used only to transcribe you more accurately, and you can see and remove any of it under “Words I use”.
• Spoken replies are generated from text. When an agent reads a reply aloud, we send that reply’s text to OpenAI to synthesize the audio. Nothing about you is sent that isn’t already in the reply, the audio is played once and not stored, and you can turn reading aloud off at any time.
• Blood-test files are kept only until they are read. When you add blood work, each PDF or photo goes to storage only your account can read, is read once to pull out the values, and is deleted as soon as it has been read. The values are shown to you to check before anything is saved; values you have not saved yet are deleted when you discard them, or after 7 days. Only the values you confirm are kept, and you can correct or delete any saved test at any time.
• We do not record video.

You, the AI agents you talk to, and any outside assistant you connect yourself, which sees only what you allowed it (see “Connected assistants” below). Every agent in your team shares one picture of you. That is the point of the product, and it is why the nutritionist knows what your labs said. No other user can see any of it; the database enforces this per row, not just in the app.

Nothing is sent to your doctor, your insurer, or any clinician. Nothing in Perfect Human is filed as a medical record anywhere. If you want your clinician to see something, you export or print it and take it yourself.

A named administrator can see account-level information for support and abuse handling.

In My Family you can keep notes about the people you look out for: relatives, friends and others. What you add is private to you. We never contact them, and nothing about them is sent to an outside assistant you connect.

It is kept apart from your own record: a parent’s blood test is read on their page and never becomes yours. Family history you add (what a relative had, and when) is used for your own advice, as the guidelines use it.

A note or care sheet leaves only when you send it yourself: as a private link that expires and that you can turn off, or as text you copy. Forget someone and everything about them goes: their page, what your team learned about them, the chats about them and their files. You can download all of it at any time. Notes about a child are kept by their parent; children do not have accounts here.

You can connect an AI assistant you already use, such as Claude, ChatGPT, Grok or OpenClaw, so it can ask your agents before it books, buys or plans. Nothing is connected until you approve it, on a screen that lists what it will and will not see.

By default an assistant sees only your everyday details: preferences, food, allergies, goals, training, sleep timing, today’s plan, supplements, trips and the rules you write for your agents. Medications, conditions, lab results, mental health, sexual and reproductive health, substance use, your chats and anything about other people are not shared this way. Choose “Advice only” and it gets your agents’ advice and none of your data.

Where approvals are switched on, an assistant you have set to “Full team” may ask to see something more sensitive, such as your medications. Nothing opens until you approve that one request inside Perfect Human with Face ID or another passkey, for a single use, a day, or a trip. Assistants you run on your own computer or server can never ask.

What you share goes to the company that runs the assistant (Anthropic for Claude, OpenAI for ChatGPT, xAI for Grok), or to your own setup if you run it yourself, and is handled under their terms, not ours. Deleting something in Perfect Human cannot recall what an assistant has already received. Memory kept only in the private vault on your device never reaches our servers, so no assistant can see it.

Connected assistants, under Account, shows a receipt for every time an assistant read something (never the content itself), and disconnects any assistant, or all of them, in one tap. Anything an assistant sends you shows up there too.

• Delete any conversation or group from within the app. Deleting removes the messages and what your agents can recall from them; choose “Delete and forget” to also remove what they learned about you from it (you see the list, and can keep any of it). Deleting a single message also removes what was learned from that message.
• Delete uploaded documents and lab reports at any time.
• Opt out of usage analytics in your account settings.
• Download a copy of everything your account holds, as one file, from Privacy & encryption in your account.
• Delete your account and everything in it from the same screen: your chats and what your agents remember, your logs, your uploaded files and recordings, your subscription, and every assistant you connected, which is disconnected first. Memory kept only on your device is not on our servers; remove it from that device. You can also ask us by emailing the address below, and we will do it.

Perfect Human is for adults. It is not designed or intended for anyone under 18, and the agents are instructed to stop and refer if they learn a user is a minor.

If we change how any of this works, we will update this page and the date at the top. For anything privacy-related, including deletion requests, contact privacy@perfecthuman.io.