Privacy policy
Last updated 13 November 2026
Three-Inbox is a private email and task workspace that I, Dr Nick Azizi, built for my own use and operate through PersisMed Pty Ltd. It is not offered to the public, and only my own Google account can sign in to it. It works with my three Gmail mailboxes, which include correspondence about patients. This page explains what the application does with that mail and the other data it holds, who else receives any of it and where, and how long it is kept.
Who receives data, and where
- Google (United States): Gmail, Google sign-in and, where I have connected it, Google Calendar. My mail stays in Gmail under my Google Account settings.
- Cloudflare (a global network): hosts the application, its database and its file storage. Requests to the application may be handled, and data held, outside Australia.
- Anthropic (United States): receives email text and other content for the AI features described below, through its API, only while AI processing is switched on in Settings. When I turn on the Web switch for a question to Ask Claude, Anthropic’s web search service also receives the words Claude searches for, and Anthropic opens pages for it only on the websites I allow.
- Claude connections (Anthropic, United States): when I use Claude at claude.ai or in the Claude apps with a Three-Inbox connection I have approved, what Claude reads through that connection goes into my Claude chats. Those chats are governed by claude.ai’s terms and my claude.ai account settings, not by the API terms that cover the AI features inside the application. The claude.ai setting that allows my chats to be used to improve Claude is off.
- My scheduled Claude task (Anthropic, United States): when I press Publish in News → Website news, the stories I picked (their ids, headlines and source links) and my note for the writer are sent through Anthropic’s routine trigger to start my own scheduled task in my claude.ai account, which writes and publishes those stories on my public website. Nothing else from the application goes with them, and the trigger’s token stays on the server.
- Push services (Apple, Google, Mozilla or Microsoft): carry the notifications I turn on to my own devices, encrypted so that the push service cannot read them.
- Others, described below: the web servers of senders whose emails contain pictures or whose logo is shown, the unsubscribe address of a mailing list I unsubscribe from, Cloudflare’s DNS service, Healthchecks.io (only if I set up the outside monitor) and Frankfurter (currency rates). None of these receives the text of an email.
Google sign-in and the Google permissions requested
- Sign-in asks Google for my name, email address and Google account id (
openid email profile). Only my own address can sign in, Google must have verified it, and the first account id that signs in is recorded so that a different Google account with the same address is refused. A signed-in session lasts up to 30 days, and Settings has “Sign out all devices”. - Gmail (
gmail.modify, for each connected mailbox): used to read messages and attachments, to change labels (read or unread, starred, archived, Spam, Trash), to create and delete drafts, and to send mail I have written and confirmed. Apart from a draft I discard, which Gmail deletes outright, the application never deletes mail permanently by itself: it moves mail to Trash. - Gmail, Delete forever (
https://mail.google.com/, only for a mailbox where I have allowed it): asked for only when I first press Delete forever in that mailbox’s Trash. It is used for one thing: deleting for good the messages in Trash of the conversations I choose, after I confirm. Nothing else in the application uses it, and Claude connections and Ask Claude cannot delete mail for good. - Google Calendar (optional, for each mailbox):
calendar.calendarlist.readonly,calendar.events.readonlyandcalendar.app.created. These let the application read my calendars and write only to calendars the application itself creates. - Google Contacts (
contacts.readonly, only for a mailbox where I have allowed it): asked for only when I first press Photos from Google in Contacts and choose Continue to Google. It is read only and used for one thing: matching the photos on my saved Google contacts to people already in my contact books (see Contact photos below). Nothing in Google Contacts is changed, and Claude connections and Ask Claude cannot read it.
How Gmail data is used
- Reading mail. Messages are fetched from Gmail when I open a folder or a conversation. A scheduler also checks each mailbox every minute for new mail (sender, subject and labels, and the whole message only when one of my rules needs it), to apply my rules and send the notifications I have chosen.
- Sending. Mail is sent only when I press Send or Send now, or schedule a message, myself, in a signed-in browser. A message I schedule, or one held for a few seconds so that I can undo it, is sent by the application’s scheduler at that time. Nothing is ever sent or forwarded automatically, and neither the AI features nor Claude connections can send.
- My rules. Rules I create in Settings can label, archive, mark read or move incoming mail to Trash. They are my own instructions and involve no AI.
- The split inbox sorts my Inbox into sections by fixed rules, without AI.
The AI features (Anthropic)
AI processing is off until I switch it on in Settings. While it is on, these features send content to Anthropic’s API. Most requests also carry the short “About me” note I write in Settings.
- Email review into Tasks (twice a day, and when I press Review email now): each new or changed conversation, other than bulk mail and automatic notices, is sent (sender, recipients, date, subject and the text of each message, up to 20,000 characters a message) with the names of its attachments, but not the attachments themselves. Claude returns tasks (one about a patient goes to Practice → Patient Tasks), entries for emails about a patient’s result (Practice → Patient Results, kept encrypted) and “Needs action” findings. The scheduled reviews use Anthropic’s Message Batches service, which holds the requests and answers until the application collects them; the application then asks Anthropic to delete each batch.
- Summarise and suggest a reply: the conversation (up to about 70,000 characters, newest first) and, when it asks to meet, my free times from my calendar.
- Reply with Claude (Claude writes the first draft of a reply when I ask): the conversation (its newest messages whole, older ones shortened), the names of its attachments, the mailbox the reply goes from, any instruction or quick reply I choose and, when it asks to meet, my free times from my calendar. While “Match my writing style from sent emails” is on in Settings (it is on unless I switch it off), up to three of my own recent sent replies from that mailbox, read from Gmail for that request with their quoted history and signature taken out, are sent too, for tone only; they are not kept. The draft only goes into my reply for me to check; nothing is sent.
- Learning from my edits (every hour, while AI is on): for each email or letter Claude drafted that I changed before sending, Claude’s first draft, what I sent and the changes I asked Claude for. Claude answers with suggested lessons about how I like things written; a lesson that names a patient, reads like a date of birth or a Medicare number, or carries a name or number from the letter is dropped before it is kept. Nothing changes how Claude writes until I approve a lesson. The lessons I approve (never one I reject or switch off) are sent with Reply with Claude, Summarise’s suggested reply and Ask Claude’s drafts, and given to a Claude connection when it writes a document or email for me. After I approve the final version of a letter or report, the lessons I approved from it are sent so that Claude can word an addition to my writer skill for that kind of document; the wording is checked the same way, nothing in a skill changes unless I approve the exact words, and the application never changes a skill itself: a Claude connection reads the wording I approved and prepares the updated skill for me to install.
- Editing a draft with Claude: the draft text or the part I select, with the text around it.
- Plain-English search: only the words I type.
- Translate (only when I press Translate on an email): that one email’s text, up to about 16,000 characters. The translation is shown to me and not kept.
- The morning brief (8 am, or when I rebuild it): my open task titles with the names of the people they are about, recent emails’ subjects, senders and short previews, findings, and today’s calendar event titles, times and places. With AI off, the brief is put together by fixed rules instead.
- Add to calendar: the conversation’s text, when it is not a calendar invitation (invitations are read without AI).
- Follow-up drafts: when someone has not replied to a conversation I asked to follow up, the conversation’s text, so Claude can word the draft. With AI off, a template is used.
- Ask Claude (the assistant inside the application): my questions and whatever its tools read to answer them, which can include whole emails, tasks, findings, the morning brief, my calendar, reminders, my contact book, my private contacts when I ask about them, and a PDF, image, Word or text attachment I ask it to read. While Finance is shown in Settings, it can also read my finance figures, including the bank transaction lines behind my recurring charges, which are then sent to Anthropic. Since 17 October 2026 that includes my personal spending lines (Personal) and the totals added up from them.
- Ask Claude’s changes: when I ask, it can propose changes to my tasks, mail (labels, archiving, Trash), reminders, findings, events on the application’s own calendars, my contact book and my finance decisions. A proposal changes nothing: it is shown to me with the words of my request, and only my Confirm applies it. It cannot send, forward or schedule mail, invite or notify anyone, change my own calendars or publish finance figures. Settings → Claude can limit it to proposing Day Sheet changes only.
- Ask Claude and the web: only for a question I ask with the Web switch on (off by default, and only while “Use the web” is allowed in Settings → Claude). Claude may search the web at most five times for that question, with words it chooses, and is told never to search for a patient or for anything from my mail. It may open pages only on the websites listed in Settings → Claude (public reference sites such as health.gov.au and ato.gov.au, until I change the list). The answer shows every search, word for word, and the pages used.
- Reading a PDF for a Claude connection: when a Claude connection opens a PDF attachment, the whole document is sent to Anthropic to be transcribed.
Some requests ask Anthropic to keep a cached copy of the prompt for a few minutes, so that a follow-up request costs less. Beyond that and the message batches, the application asks Anthropic to keep nothing. How long Anthropic keeps what it receives, and what else it may do with it, is governed by Anthropic’s commercial terms for my API account. The application itself does not use any of this data to develop, improve or train AI models.
What the AI switch stops, and what it does not. Switching AI processing off stops every feature in the list above, including PDF transcription for Claude connections, and review batches still waiting at Anthropic are cancelled by the scheduler. It does not stop Publish in Website news, which starts my own scheduled Claude task when I press it. It does not stop Claude connections: a connection I have approved can still read my mail through claude.ai with AI processing off, because that is me using Claude, not the application calling Anthropic. It also does not affect features that use no AI: finance capture, file exports, my rules, notifications and the rest of the application.
Claude connections
- I can connect Claude (claude.ai and the Claude apps) to Three-Inbox. Each connection needs me to sign in and approve it, either for reading only or for reading and changing, and the application notifies me, on the devices where I have turned notifications on, whenever a new one is approved. Settings → Claude lists each connection and can disconnect it.
- A connection can search and read my mail in full (including attachment names), and read my tasks, task packs, morning brief, findings, reminders, scheduled emails, calendar, contact book, private contacts and the system status. A connection allowed to change things can also read attachments, list and save Gmail drafts and labels, label, archive or move conversations to Spam or Trash, edit tasks, contacts (private contacts only after asking me) and reminders, add events to the application’s own calendars, create download links to original files (below), and upload files (such as a letter Claude wrote) to attach to the drafts it saves. It cannot send or forward mail, and apart from discarding a Gmail draft, nothing it does deletes anything permanently.
- In Settings → Claude I can limit which mailboxes Claude connections can reach: all three, Personal only, or Personal and PersisMed. A mailbox left out is off-limits to every connection: its mail, drafts, findings, reminders and scheduled emails, and the tasks that came from its email, are neither read nor changed, and nor are the calendar events the application made for them. If that setting cannot be read, a connection is limited to Personal until it can. The assistant inside the application is not affected by this setting.
- A connection gets my finance figures only if I tick a box for that when I approve it: one box to read them (including the bank transaction lines behind my recurring charges), and one, meant for my finance engine on my Mac, to publish them. Without a tick a connection can neither read nor publish them. The mailbox limit above applies to the receipts among them. Since 17 October 2026 reading them includes my personal spending lines (Personal), and a connection with the publishing box can also correct how a spending line is sorted, which it does only when I ask and which is recorded in Changes by Claude. Since 13 November 2026 a connection with the publishing box can also ask my finance engine on my Mac to refresh or publish the figures, which is recorded in Changes by Claude; it never holds the engine’s own key.
- The changes a connection makes to my tasks and task packs, to mail (labels, archiving, Spam and Trash), to findings, and to contacts’ existing details are recorded in the application so that I can review them, as are the changes I confirm in Ask Claude.
What the application stores
The application stores the following in its Cloudflare database and file storage. Items marked encrypted are encrypted by the application (AES-GCM) with a key held as a Cloudflare secret, before they are stored.
- Google access: each mailbox’s Google tokens, encrypted.
- Sent emails: a copy of each email I send from the application, exactly as sent (recipients, subject, text and attachment names), encrypted, kept for my records.
- People I have emailed: each address an email I send from the application goes to (To, Cc and Bcc), with the name it was written with, when I last emailed it and how many times, encrypted, so the address fields can suggest it next time. Never the subject or the text. Received mail adds nothing, and nobody is added to my contact book.
- Drafts I am writing, including drafts Claude prepared and what Claude based them on, and the files I attach to them. These are not encrypted by the application.
- Finished paperwork documents that a Claude connection wrote for a task and kept with it, such as a referral letter, which are about a patient. The file is kept in the application’s file storage, not encrypted by the application; its name, size and checksum are kept, encrypted, with the task. I can open, download or delete it from Tasks, and a connection allowed to change things can attach it to a draft or get a 15-minute download link for it, which is recorded like any export.
- Learning from my edits: when I send an email Claude drafted, or a task’s Word letter or report that Claude wrote first, the application keeps Claude’s first draft beside what I sent, encrypted, so that Claude can learn how I like things written. The words are kept for 30 days after an email is sent and 90 days after a letter or report is sent; after that only the numbers stay (how many words Claude wrote and I sent, the share I changed and how many facts I changed). The copy of the sent email itself is unchanged and never holds Claude’s draft. When I ask Claude to change a letter it wrote, what I asked for is kept with the task too, in general words and never with a patient’s name, date of birth or Medicare number. The lessons Claude suggests from my edits, and my decisions on them, are kept, encrypted, until I remove them; a lesson I reject is kept so that it is not suggested again. When I approve the final version of a letter or report, the application notes when, and the skill wording Claude suggests from it, with my decision on it, is kept, encrypted, the same way.
- Files a Claude connection uploads to attach to a draft, such as a referral letter Claude wrote, which can be about a patient. They are kept in the application’s file storage, not encrypted by the application, for 24 hours and then deleted; a draft they were attached to keeps its own copy, like any draft attachment. Each upload is recorded (file name, size, checksum, which connection and when), encrypted, never the file’s content. A link to upload a larger file works once, for 15 minutes, without signing in.
- Tasks, task packs and the record of Claude’s changes to them and to my mail, reminders, findings, calendar, contacts and finance decisions (which items and what was done), encrypted. A task taken from an email keeps the sender, subject, date and the sentence it was taken from, and may name a patient.
- Needs action findings (Claude’s note of what an email needs, with its quote and any suggested reply), encrypted; the dates and who closed one are kept readable so the application can sort them.
- Morning briefs, encrypted.
- Reminders (snoozes, follow-ups, Reply Later), encrypted.
- My rules and the record of what they did, encrypted.
- Notification devices (each device’s push address and keys) and waiting notifications, encrypted.
- Calendar: events read from Google Calendar (titles, times, places), encrypted; each copy is used for at most 30 minutes.
- My own calendar: events I keep in the application itself rather than in Google Calendar (title, times, place, notes, a web link if I add one, how it repeats, and the one occurrence changed or removed), encrypted, apart from the first and last day each can fall on. They are kept until I delete them; a deleted one is kept so Undo works. They are never sent to Google.
- Remembered AI answers: the last summary of a conversation for 7 days, and a PDF transcript made for a Claude connection for 1 day, both encrypted.
- Ask Claude’s proposals waiting for my Confirm, for 30 minutes, and what is needed to undo a change I confirmed, for 7 days, both encrypted. Ask Claude’s conversations themselves are not stored.
- Subscriptions: the list of senders whose emails offer unsubscribing (each sender’s name and address, the subject of their newest email, how many there were and when), encrypted, kept for 6 hours so that the list opens quickly; and the addresses of the senders I have unsubscribed from, with the date, not encrypted by the application, kept until I mark a sender as still subscribed (at most the latest 500).
- Message details cache: sender, recipients, subject, date, labels and the short preview Gmail gives, for each message in a list I open, so that the inbox loads quickly. This is not encrypted by the application. Each entry is used for at most 24 hours, and is deleted the next time that mailbox’s list loads or when its mail changes.
- My contact book of clinicians, practices and services (names, specialties, practice addresses, phone, fax and email, notes on what they do), not encrypted by the application. It is meant for clinicians and services, never patients, and no automatic process adds anyone to it. When I send an email from the application to an address in the book, that contact is marked as written to by email on that date; nothing about the email is recorded. When a referral or other document addressed to a contact in the book reaches the drafted stage, its kind (for example “Referral letter”) is recorded against the contact, never the patient.
- My private contacts, a second contact book apart from the one above, for friends and family, health professionals (doctors and allied health), patients, personal advisers, trades and the like (names, relationships, household, phone numbers, email and postal addresses, birthdays and anniversaries, business details and my notes), encrypted. Only I move or remove them, and a Claude connection adds or changes one only after it has asked me, or when I have asked it to find that contact’s details online and save them; when I import them from a contacts file, the file is read on my device and only the contacts I choose, one at a time, are saved. When I send an email from the application to one of them, the contact is marked as emailed on that date, and how many times; nothing about the email is recorded. Referrals and other paperwork are never addressed from it. A contact I remove is kept, hidden, so that I can undo it.
- Saved receipts and exports (below): the original files and a record of each (sender, subject, file names), not encrypted by the application.
- Finance figures that my finance engine publishes (account balances, GST drafts, deadlines, receipts it found in my mailboxes, ideas, its briefing, and my recurring charges with the bank transaction lines each was found in: date, amount, the bank’s description and which account; all of them proposed figures; since 17 October 2026 also my personal spending lines from my personal card and everyday account, for Personal: merchant, amount, date, category and which account, as my finance engine sorted them, never the home loan), my decisions on them (including my corrections to how a spending line is sorted, since 17 October 2026), the extra mailbox sweeps I ask for and the questions I put to Claude’s weekly research, encrypted. Which mailbox a receipt came from and the dates the figures are read to are kept readable, so the application can sort them and apply the mailbox limit. A record of each publish (which connection, how many documents, any problems found; never a figure) is kept too, not encrypted.
- The finance engine’s jobs (since 13 November 2026): what I ask my finance engine on my Mac to do from Finance → Engine, or a connection with the publishing box asks it to do (refresh the figures, publish, or file a statement), with each step’s progress, any question it asks me and my answer, and its result line, encrypted; never a figure (anything that looks like an amount is masked before it is stored). What kind of job it is, its status, who asked and the times are kept readable. The engine’s worker on my Mac also reports, every few seconds, when it last checked in, whether the finance drive is connected, the engine’s version and my Mac’s name.
- Bank and card statements I add on Finance → Engine (CSV exports only, since 13 November 2026): kept in the application’s file storage, encrypted, only until my Mac has filed them, and handed to my Mac by a link that works for 15 minutes. They are never patient or clinical files.
- Sender logos, kept by the sender’s domain name alone (for example racgp.org.au).
- My settings, including my signatures and my “About me” note.
- Records the application keeps about itself: how much each AI feature was used each day (counts of calls and tokens, never content), which conversations the email review has handled (by id), the state of its own health checks, the daily news briefing (public news, nothing from my mail), and the Website news candidate stories for my public website (public news, shown for 7 days and kept 120 days so the daily task can learn which kinds of story I pick) with a record of each time I press Publish (kept 12 months).
- Claude connections: each connection’s name and approval, and its access tokens stored only as one-way hashes.
Original files and saved receipts
- Exports. A Claude connection allowed to change things can copy an email’s original attachments (and the whole email, if asked) into the application’s file storage and return a download link. The link works for 15 minutes for whoever holds it, without signing in. The files are deleted after 7 days. Each export is recorded (mailbox, message, file names and checksums, and which connection asked), and those records are deleted after 12 months. No AI is involved.
- Saving receipts and invoices automatically is off unless I switch it on. When on, every hour the application looks at the last two days’ received mail (or further back, up to 30 days, when it has fallen behind), by sender, subject, preview and labels, picks out receipts, invoices, statements and bills by fixed rules, and keeps a copy of their attachments and the email itself. In my Professional and PersisMed mailboxes it keeps only email from a fixed list of finance senders (the tax office and business registers, my bank and card, accounting software, software subscriptions and suppliers). In every mailbox, the rules leave out email from clinical senders (health services, pathology, health funds, HICAPS, Tyro Health and Medipass), email with patient, referral, pathology, discharge-summary or date-of-birth wording, and email whose text has a patient’s details (a Medicare or IHI number, a date of birth, “patient” or “consult for”), checked once the email is read in full; in my Professional and PersisMed mailboxes they also leave out any clinical wording, government health senders and anyone in my contact book. Anything left out is not saved; I can still export it myself. No AI is involved. A saved copy is kept until I mark it picked up, and deleted 90 days after that.
Google Calendar
- If I connect a calendar, the application reads my calendars’ event titles, times and places, and whether I declined an event. It does not keep event descriptions or guest lists. Holidays in Australia is read by default.
- Bringing a calendar in: when I press Bring in for one of my own Google calendars (Settings → Calendar), the application reads that calendar’s events from the past year and everything ahead (title, times, place, description, reminders, how each repeats and the occurrences moved or cancelled) and keeps them in my own calendar, encrypted, the description as the event’s notes. Guests, attachments and video links are not kept, and events I declined are left out. Nothing in Google Calendar changes and nobody is told. Once a calendar is brought in, the application shows it from my own calendar instead of reading it from Google. Check against Google reads that calendar again to show me what is new, changed or removed there; it changes nothing until I press to bring those in or remove them here.
- Disconnect Google Calendar (Settings → Calendar): when I press it, the application stops reading and writing Google Calendar on every account, as Remove Calendar does for one; my own calendar keeps everything. Gmail stays connected and nothing in Google is deleted.
- Inviting people to an event: Invite… in an event’s panel saves an email draft in my Drafts, from the mailbox I choose, with an invitation file attached (the event’s title, time, place and link, never its notes). Nothing is sent and nobody is told until I open the draft and send it myself.
- Backup of my calendar: Download calendar (.ics) makes a file of my own calendar’s events (title, times, place, notes, repeats, reminders) on the device I press it on. It is not kept or sent anywhere else.
- Events I add in the application, events for my tasks (if I choose that in Settings) and events Claude adds when I ask are kept in the application itself, in my own calendar (see “My own calendar” above), not in Google Calendar. The application writes to Google Calendar only to change or remove an event it made there before, on one of the two calendars it created, “Three-Inbox · Work” and “Three-Inbox · Personal”. No event invites or notifies anyone. An event for a task names a patient by initials unless I choose otherwise in Settings.
- My free times can be put into a reply I am drafting, and sent to Anthropic for Summarise as described above.
Notifications
If I turn on notifications for a device, each notification is encrypted for that device before it leaves the application (RFC 8291), so the push service cannot read it. What it says depends on its kind and on the lock-screen level I choose in Settings: by default a reminder shows the sender and subject, a new-task notification shows the first task’s title, and a new-email notification shows the sender and mailbox, with the subject only if I turn that on. A Claude connection allowed to change things can send a short message of its own when an unattended run finishes something I asked for or finds something urgent: at the “sender and subject” level its own words show, at the lower levels only that Claude has an update; it is refused if it names a person in my tasks, a date of birth or a Medicare-like number, at most 5 an hour and 20 a day are allowed, and I can switch these messages off. A finance notification uses fixed words, such as “Finance: something is due in 3 days”, and never an amount, an account or a payee; the same goes for the two about my finance engine’s jobs (“Finance: a job on your Mac failed”, “Finance: your Mac needs an answer”). A notification never contains an email’s body, a quote from it or an attachment. Waiting notifications are held encrypted and deleted within a week. A reminder for an event on my own calendar comes at its time even in quiet hours; at the generic level it shows only the time, at the sender-and-subject level the event’s title and time, and at the full level its place too. A reminder for a task’s time on the calendar shows its title (which names the patient as I choose in Settings → Calendar) at every level, the generic one included.
Pictures in emails, sender logos and other outside requests
- Pictures in emails load by default; I can block them for a message. They are fetched by the application’s server, not by my browser, without cookies or a referrer, so the sender’s server sees a request from Cloudflare rather than from my device. It can still tell that the email was opened. Cloudflare’s DNS service is asked for the picture’s server address.
- Sender logos: to show an organisation’s logo, the application’s server asks Cloudflare’s DNS service for the domain’s published logo record (BIMI), fetches the logo that record names (which may be on another website), or asks the organisation’s own website for its icons. Only the domain name is used, never an address or any message content, and my browser never contacts the site. Senders at mailbox providers such as Gmail or Outlook are never looked up. A stored logo that has not been looked up again for about seven months is deleted.
- Unsubscribe: when I press Unsubscribe on an email, or on the senders I choose in Subscriptions, and confirm it, for each email that offers one-click unsubscribing, the application’s server sends the unsubscribe address written in that email’s own headers (as Gmail holds them) a fixed request, “List-Unsubscribe=One-Click”, without cookies or a referrer; only public web servers are contacted. Before unsubscribing from senders in Subscriptions, the application reads only their emails’ sender and unsubscribe headers from Gmail and shows me who it will ask. If an email asks for an email instead, a new message opens for me to send myself; if it names a website, that page opens in my browser. When I open Subscriptions, the application reads from Gmail only the sender, subject, date and unsubscribe headers of the mailing-list emails of the last 90 days (at most the newest 400 in each mailbox) and lists the senders; nothing in the mailbox changes and nothing is sent until I press Unsubscribe. Claude connections and Ask Claude cannot unsubscribe.
- Healthchecks.io, only if I set up the outside monitor, receives a bare request every five minutes while the application is answering, so it can alert me when it stops. It carries nothing about my mail, contacts or tasks.
- Frankfurter (api.frankfurter.dev) supplies the European Central Bank’s US dollar to Australian dollar rate, at most twice a day, so that AI costs can be shown in Australian dollars. The request names the two currencies and nothing else.
- Find details online on a contact sends nothing from the application: it opens a new chat in my own Claude account with my add-contact request and that contact’s name already typed in (“/add-contact” and the name), and I choose whether to send it. It carries only that contact’s name from my contact book. On one of my private contacts it carries only the contact’s name and business name, with a request to find their public professional or business details and save them to that contact; the details Claude finds are then saved without asking me again, and it tells me what it saved and where each came from. In both books the request also asks Claude to check the person’s LinkedIn and to look for a photo of them, in this order: their own or their firm’s website (and, for a health professional, hospital and directory profiles such as Healthshare, HotDoc and HealthEngine), then LinkedIn, then their social media (Facebook, Instagram, X and any other public profile), passing it on only if it is clearly them. Find photo online works the same way but asks only for a photo: it carries the contact’s name and where they work (for a work contact, their specialty and main practice; for a private one, their business name), and nothing else about them is looked up or changed.
- Contact photos: a contact in either book may have a photo, shown in place of their initials. Either I choose one from my own device (the page makes it smaller before it is sent), or I take them from the photos on my own contact cards (Photos from iPhone, and Import, which saves a card’s photo with the contact I add: the file I export is read on my device, never uploaded, and only the photos I choose are saved), or from my saved Google contacts (Photos from Google: the application’s server reads their names, emails, numbers and photo addresses with the read-only permission above, keeps none of it, and shows me only the photos of people already in the book, which load through the same checks as pictures in an email; only the photos I choose are saved, and nobody is told), or from Instagram through my own script on my iPhone (Photos from Instagram: I put a private contact’s Instagram account on their contact; the script runs only on my phone, opens their Instagram page logged in as me on Instagram’s own page, never sees my Instagram password, and sends the profile picture here with a key I make and can turn off; the key, kept only as a hash for at most a year, can list the private contacts that have an Instagram account, find one by name, add an account and save a photo, nothing else, and never replaces a photo I chose; this application never contacts Instagram), or Claude gives the address of a public picture of them and the application’s server fetches it, through the same checks as pictures in an email (a public site only, a picture and nothing else, at most 2 MB). The photo is stored encrypted, with only who added it, when, and the name of the site it came from. A photo I chose is never replaced by Claude’s, and only I remove one; removing a contact keeps its photo so that Undo brings both back.
- Cloudflare may keep records of requests to the application under its own terms. Some request addresses contain a picture’s web address from an email, an exported file’s name or a sender’s domain. Searches are sent inside requests rather than in their addresses.
Google API Services User Data Policy
Three-Inbox’s use and transfer of information received from Google APIs to any other app will adhere to the Google API Services User Data Policy, including the Limited Use requirements. Gmail and Google Calendar data are used only to provide the features described on this page, for me. They are not sold, not used for advertising, and not used by the application to develop, improve or train AI or machine-learning models. They are transferred only as this page describes: to Anthropic for the AI features I switch on; to Claude, through a connection I have approved, when I use it; to Cloudflare, which hosts the application; to push services, encrypted, as notifications I turn on; to whoever holds a download link I asked for; to a sender’s web server, as a request for a picture in that sender’s email; to websites that publish a sender’s logo, as a request for that logo, using only the sender’s domain name; to a mailing list’s own unsubscribe address, when I unsubscribe from it; and to Google Calendar, as events I add. Apart from me, no person reads this data through the application.
How long data is kept, and deleting it
- A sent email’s copy is kept until I move its conversation to Trash in the application and confirm that the saved copy should go too. Nothing deletes it automatically.
- An address remembered for suggestions is kept until I choose Forget on its suggestion; emailing it again remembers it again.
- A draft stays until I send or discard it; its attached files are deleted when it is sent or discarded.
- A file a Claude connection uploads is deleted after 24 hours.
- Claude’s first draft and what I sent, kept for learning from my edits, are deleted 30 days after an email was sent and 90 days after a letter or report was sent; the numbers about how much I changed are kept.
- A finished paperwork document kept with a task is deleted 90 days after that task is done or removed (unless it is restored), or when I delete it.
- Tasks, task packs, contacts and reminders are kept; a task or contact I remove is kept, hidden, so that it can be restored. The record of Claude’s changes is kept. A proposal I do not confirm is deleted after 30 minutes, and the means to undo a confirmed change after 7 days.
- A finding stays while it is open, and is deleted 12 months after it was closed. Morning briefs are deleted after 30 days.
- Records of the email review’s runs, of exports and of what my rules changed are deleted after 12 months; the list of what my rules did lately, after 30 days.
- A statement I add on Finance → Engine is deleted once my Mac has filed it (or found it already filed), and in any case after 7 days. A finance engine job is deleted 90 days after it finished.
- Finance figures are kept until I remove them. A figure my finance engine replaces keeps the version it replaced beside the new one, and a document it withdraws is hidden but kept. My decisions on them and the record of publishes are kept; nothing deletes them automatically.
- Exported files are deleted after 7 days. Saved receipts are kept until I mark them picked up, and deleted 90 days after that.
- Disconnecting a mailbox in Settings deletes its Google tokens, its cached message details and calendar events and its new-mail markers, and asks Google to revoke the application’s access. It does not delete the drafts, sent copies, tasks, findings, reminders or saved receipts that came from that mailbox. I can also revoke access at any time from my Google Account permissions.
- The application has no single “delete everything” action. Deleting its Cloudflare database and file storage removes what the application stores. It does not remove what others hold under their own terms: requests Anthropic received, my Claude chats, my mail and calendar at Google, and Cloudflare’s request records.
Contact
Questions about this policy: nickazizi@persismed.com.au.