When what you say needs to stay private and secure.
Haven encrypts message bodies, file contents, filenames and the names of new groups in your browser before they leave it, and pads message plaintext up to a fixed size bucket, so a stored ciphertext gives away a range rather than a length. The server holds ciphertext and wrapped keys that are inert without a password we never see. Signing up takes a username and a password: no email, no phone number, nothing already attached to your legal name.
What we don't do, and why it helps you. We don't ask for an email address or a phone number, so nothing at sign-up ties your account to your name. We don't keep logs of your IP address, so there is no record of where you connected from for anyone to ask us for. And we don't run analytics or trackers, on this site or in the app. The IP part is a setting on our server, which this page cannot prove.
- No phone number, no email
- Messages and files encrypted in your browser
- TweetNaCl + Ed25519
- Calls relayed by default: the other person does not get your IP, our relay does
- No analytics, no trackers
- No IP logs: your address is not written down, a setting on our server this page cannot prove
Free tier: no card, no email, nothing at signup already tied to your name — and the same metadata as every other tier. Paid plans start at $14.97 a month and can be bought with a card or with a voucher that attaches no name. Like every encrypted app that runs in a browser, Haven relies on the server that delivers its code, and with hosted Haven that server is ours. The threat model explains what that means for you, and is worth reading before you sign up.
Haven for Android
A signed APK you install yourself, on Android 7.0 or later. It needs no Google Play services and carries no tracking SDK. It does not yet raise a notification when a message arrives.
Scan it with your phone's camera to download the APK there.
5a0cbf2a4a4e87c14d1b1a380e10fab3c49a9e022c15205f45d971d24f2da624
A messenger first. The checks stay in view.
Conversations, files, voice notes and calls, in an interface that gets out of the way. What makes it Haven sits in plain sight above every conversation: whether the key log matched, and whether anything has been removed from the transcript since you last read it.
The window beside this is drawn from the app's own colours and type, not a mock-up from a design tool. Open the app and it looks the same.
A managed messenger, with its limits in plain view
Haven is an end-to-end encrypted messenger, sold as a managed service. A subscription buys the part that is work: the machine, the backups, the upgrades, the storage, the push delivery, and someone whose job it is when any of that breaks. It does not change the cryptography — that is the same code at every tier, including the free one — and it does not remove the trusts this design cannot encrypt away.
Every security tool has limits, and the right choice depends on who you need protection from. Here are Haven's, in plain words. If someone well resourced is looking for you in particular, the five boxes below are the ones to weigh first; they are here on the homepage so you can decide with everything in front of you.
The app's code comes from our server
A web app has no code signing. Each time the page loads, your browser fetches the app's JavaScript from the server, with no signature and no version pinning. In principle, a server could hand one person a modified copy and everyone else the normal one, and that person would have no way to tell. End-to-end encryption can't guard against this, because the encrypting is done by the code the server sent.
With hosted Haven, that server is ours. Proton, Tuta and every other encrypted service that runs in a browser share the same limit, and none has solved it yet. If you need protection from someone who could force a hosting company to act, this is the limit that matters most, and the threat model explains it in full.
The server knows who talks to whom
Conversation membership is stored, not encrypted, in an ordinary database table, and one query can list every pair of accounts that talk to each other. The server's message timestamps are kept to the hour, and in order, and the sender's signed time is to the second or finer on messages from apps not updated since 17 November 2026. Haven protects what you say, not the fact that you said it. If the fact of a conversation is the sensitive part for you, SimpleX or Session hide it better today.
Forward secrecy is optional, and partial
There is an optional per-message ratchet, off unless a conversation turns it on: each message then gets its own key, and your device destroys each message key 90 days after it first reads or sends it. It is a hash ratchet, not Signal's Double Ratchet, so it does not recover after a key is stolen. Separately, a conversation's key changes on a schedule, and you can ask the server, one conversation at a time, to delete its copy of the old ones. Until you do, it keeps every historical conversation key, each wrapped to your long-term identity key, so if your password were ever cracked, the whole history could be opened. Switched on, that shrinks to about five weeks, as long as the other person switches it on too. Choose a strong password: one that could be guessed in five years would still open the conversations you are having today. Every edge of it is in the threat model.
It hasn't been audited
Haven follows careful practice, but it hasn't yet had an independent professional audit. Our own review found real, exploitable bugs in shipped code, including one that let a three-person conversation pass as a private two-person chat. Rather than fix them quietly, we published them in full on the security page.
Your device is part of the security
Anyone who has your unlocked, signed-in device, and any malware running on it, can read what that device can decrypt. Auto-lock, remote wipe, the panic key and the duress password help when a laptop is borrowed or taken. They can't remove malware that is already running, and clearing browser storage is not the same as securely erasing a disk.
What is true, and checkable
- Message bodies, file contents, filenames, MIME types and new group names are encrypted in the browser with TweetNaCl, before they reach us.
- Message plaintext is padded to fixed buckets from 32 bytes to 16KB, so a stored ciphertext leaks a range and not a length.
- Registration takes a username and a password. There is no email field and no column behind one —
users.emailwas dropped from the schema — and no phone number field at all. - Hiding your presence stops the server writing your last-seen time, not just showing it: the statement that stamps it carries the preference as a clause, and switching the setting on clears whatever was already stored. It does not hide that you signed in: every sign-in writes a line to your account's security log, with the time to the second and a device label, which you can read in Settings and the instance administrator can read too. The privacy page says what else is kept about a sign-in, and for how long.
- Private keys are wrapped with your password (PBKDF2-SHA256, 600,000 iterations) and never leave the browser unwrapped — which is also why there is no password reset, and cannot be.
- 60-digit safety numbers, hashed over 5,200 iterated rounds, and a key change on a contact you had already verified genuinely blocks sending rather than showing a dismissible warning.
- Keys are pinned on first sight, in your own browser — both the identity key and the message-signing key — and a later change in either raises a warning.
- Every message carries an Ed25519 signature over its conversation, author, key epoch and ciphertext, checked against that pinned key, so we cannot move a message to a different author or a different room. Messages written before signing existed are marked unsigned rather than treated as verified.
- Every message also carries a second signature over its position — the messages its author had already seen, each named by a hash, in a payload with no ciphertext in it. Take one out of the middle of a conversation and the messages that named it point at nothing, and your app says the transcript is not complete. What that does not catch is set out in the threat model: a message suppressed at the very end, a history served short from the bottom, who emptied a message that was deleted rather than removed, and what a message actually says — content is the first signature's job, and neither signature is evidence for the other's claim. Each of those is a real gap, not an edge case.
- Settings → Your data tells you what this server holds about you, enumerated from the live database at the moment you open it rather than from a list we maintain: 55 tables, 316 columns, 46 items — 38 we can read, 7 we hold and have no key for, 1 a one-way hash. Those numbers moved when this became a paid service: the nine tables the commerce work added, including the ones that hold an invoice and a redeemed voucher, are described on that panel like everything else rather than left out of it. So are the two that invite links added, one of which records that you asked to join a group you are not in — the single new readable fact that feature cost. A column added by a migration and left unclassified makes that panel declare itself incomplete instead of quietly shortening. Alongside it: a full account export, a security check-up that never asks you to redo something, and, in each conversation's own menu, a protection status that reports "partial" rather than rounding up. The limit on all four, which the app states inside them: the code drawing them was served by the machine they describe, so deriving the list from the live schema stops it going stale and does nothing about a server that wants to lie.
- A conversation's key rotates on a schedule, and immediately when somebody leaves a group or is removed from one. Adding somebody to a group is itself a rotation, and a new member is not served anything sent before they joined. The server has never held a conversation key and cannot rotate anything itself, so it names one participant's browser to do it and refuses any second attempt at the same key version — two clients acting at once cannot fork the conversation.
- Per conversation, you can tell the server to delete its copy of each superseded key once a grace period has passed. It is off by default, it leaves the current key and anything inside the grace period untouched, and history you then lose is lost permanently — the threat model lists every condition.
- Calls are relayed by default and never fall back to a public STUN server, so placing one does not hand your IP address to the other participant or to a third party — the relay is ours, and it sees both IP addresses and a call's timing and duration, but not a second of its audio. It keeps no log of those addresses or of the relay login, which we can match to your account because our deployment guide sets up per-account relay credentials; it holds them while it relays the call, and that is a setting on the server, which this page cannot prove. What keeps the server itself out of a call is the key each browser signs its connection details with, which the server hands out: until you compare safety numbers with someone, your device trusts whatever key the server gives for that person and tells you when it changes. Calls with someone you have verified are paused, on the device where you verified them, when their key changes, until you acknowledge the change there; compare the safety numbers before you do.
- A panic key and a duress password, both of which wipe local data. The duress one then shows the sign-in screen with no message saying why, so a watcher sees a signed-out device, not an unlocked one.
- An
Onion-Locationheader, so an instance published as an onion service advertises itself and Tor Browser can move you there. - No analytics, no trackers, no third-party fonts, no CDN scripts — on this site or in the app. Not "anonymised": none.
Four steps, and where the trust sits in each
Your keys are made in your browser
A keypair is generated locally. The private key is encrypted with a key derived from your password and only then does the wrapped blob reach us, which is what lets you sign in from a fresh browser without us ever being able to unwrap it.
Every conversation gets its own key
A fresh random symmetric key per conversation, encrypted individually to each participant's public key. We store those wraps. They are inert without a password we never see.
Padded, encrypted, signed, sent
Text, filenames, MIME types and group names are padded up to the next size bucket (32 bytes through 16KB) before encryption, so a stored ciphertext gives away a range and not a length. Attachments are padded before they are encrypted, so our server sees only a size band, not the exact size. Files sent before 17 November 2026, or from an app that has not been updated since, keep their exact size. The finished message is then signed twice in your browser — once over its contents, which is what stops us moving it to another author or another conversation, and once over the messages you had already seen, which is what makes one removed from the middle of the thread visible instead of silent. Neither signature covers a message suppressed at the very end, or a history we serve you short from the bottom: every half is in the threat model.
Compare a 60-digit number out of band
Read it aloud on a call, or scan the QR code in person. Haven pins every key it has seen, in your own browser: if one changes you get a banner, and for a contact you had verified, sending genuinely stops rather than warning and continuing. That check is what turns a silently swapped key into something you can see.
A messenger, and only that
Haven does one job. It is not your intranet, your file server or your conferencing platform. Every feature below is listed with what it does not do as well as what it does, so you know exactly what you are getting.
Messaging
1:1 and group chats, replies, edits, reactions, mentions, pinned and starred messages, typing indicators, read receipts, disappearing timers, and view-once messages in one-to-one chats (text, photos, audio or video). Search runs on your device, because the server can't index what it can't read.
Files and voice notes
Attachments are encrypted client-side like everything else, with EXIF and XMP stripped from images before encryption — encryption hides a photo from us, it does nothing about the GPS coordinates the recipient decrypts along with it.
The limit worth knowing before you pay for a bigger one: a file is still encrypted and sent whole, which costs roughly three times its size in browser memory and cannot be resumed if the transfer drops. The per-file ceilings the plans quote — up to 2GB — are what the server accepts, not what a phone will manage. The server side of a chunked, resumable transfer is built and tested; the browser half is not written, so until it is, this is the ceiling that applies.
Calls
Voice and video over WebRTC, encrypted with DTLS-SRTP. Calls are relayed by default rather than connected directly, because a direct call hands each participant the other's real IP address — so we carry the packets and can see that a call happened, for how long, and both IP addresses, with the relay login (which we can match to your account, because our deployment guide sets up per-account relay credentials). The relay keeps no log of those addresses or that login: it holds them while it relays the call, and that is a setting on the server, which this page cannot prove. The relay cannot hear the call: the media is encrypted between the browsers on the call. An instance with no relay reports calls as unavailable instead of falling back to a direct connection. Group calls hold up to six people: everyone connects to everyone through the relay, so there is no media server that could listen in, and each person's connection details are signed with their key, which is what keeps the server out of the middle. The server hands out those keys, though: until you compare safety numbers with someone, your device trusts whatever key the server gives for that person and tells you when it changes. Calls with someone you have verified are paused, on the device where you verified them, when their key changes, until you acknowledge the change there; compare the safety numbers before you do.
Devices, and getting out fast
Every sign-in is listed with a device label and last-active time. Revoke one, wipe one remotely, or sign out everywhere. Auto-lock clears decrypted keys from memory after inactivity. A panic key wipes local data on a keystroke, and a duress password wipes it instead of unlocking, then shows the sign-in screen without saying why.
The label ("Firefox on Linux") is how you spot a session that isn't yours — and it is
also something the server can use to link two of your accounts, if it is an unusual one
and both appear at the same time. A preference stores 'Unknown device'
instead. Off by default, because turning it on costs you the one way you had of
recognising your own sessions.
Shared plans, and what they include
A plan that holds more than one account lets its owner add and remove people on it. That is the whole of it — no administrator here can read a message, because the content is sealed. They can see who is in a conversation and when it happened, like everyone else with database access; the threat model's first table is the full list. On Business and Enterprise the owner can also set how long the server keeps its copy of messages. That deletes sooner and still reads nothing, and it does not touch conversation membership or timestamps, which survive the sweep.
What isn't encrypted
Conversation membership, message timestamps (the server's to the hour; the sender's signed time to the second or finer on messages from apps not updated since 17 November 2026), attachment byte counts, reactions, typing indicators and per-conversation settings such as the disappearing timer are all stored in the clear. The first three matter most: together they show who talked to whom, when, and roughly how much was sent. We list them so you never have to guess.
Six plans, and the arithmetic in the open
Prices in US dollars. A yearly plan is one payment covering twelve months, and the saving is different on each plan — 20%, 23% and 25% — so each card shows its own number instead of rounding all three into one slogan.
What you are paying for is operational: a hosted instance somebody keeps patched, backed up and running. It is not a stronger guarantee than the free tools give. It also costs more than every published price in this category, and the comparison below shows by how much.
Team plans are per seat, billed from a minimum seat count.
Full comparison of all six plans →
How buying one works
A plan button hands you to the app's sign-up with that plan carried in the link. You make an account — a username and a password, no email, no phone number — and the app starts the checkout. Nothing is charged before the account exists.
Two ways to pay, and the difference matters more here than on most sites. A card, through Stripe's own page, which never lets a card number near this server and does tie your legal identity to this account — for a determined adversary that invoice is worth more than the ciphertext. Or a prepaid voucher code, which attaches no name at all: you buy it from us however you like, redeem it in the app, and nothing about the payment ever reaches a processor. An operator can always join the redemption to the account that redeemed that code. How you get one, and the exact size of that anonymity, is on the pricing page.
One consequence worth knowing before the first payment rather than after it: Haven cannot email you about your bill, because no account holds an email address. No receipt, no renewal reminder, no warning before a card expires. Stated in full here.
What else you could use — and when you should use it instead
The tools closest to what this page promises are free. Signal, Session and SimpleX cost nothing, are donation funded, and two of them hold less about you than Haven does by construction rather than by policy. If one of them fits, use it. We would rather say that than sell around it.
| Haven | Signal | Session | Telegram | |
|---|---|---|---|---|
| Needs a phone number | No: a username | Yes | No | Yes |
| Operator sees who talks to whom | Yes, stored in the clear | Mitigated (sealed sender) | No central operator | Yes, fully: retains content |
| End-to-end encrypted by default | Messages and files | Always | Yes | No: opt-in, 1:1 only |
| Forward secrecy | No: off by default; opt-in and partial, per conversation | Yes | No | Secret Chats only |
Signal, Session and Telegram as our competitor research found them in August 2026, from their own documentation; Haven as this code does it.
- Signal Free Donation funded, its protocol independently analysed, real per-message ratchet. Needs a phone number. The default recommendation for almost everybody.
- SimpleX Chat Free No account identifier at all — nothing for a server to build a graph out of. Strictly better than Haven on the metadata Haven leaks.
- Session Free · Pro tier in beta, no price published No phone number, routed over an onion network so the server never sees your IP. A paid tier is coming and has no announced price.
- Threema Private 6.00 one-time Buy it once, keep it. Bitcoin accepted specifically so the purchase can be anonymous — Haven's answer to the same problem is a prepaid voucher, bought however you and we agree.
- Threema Work 3.00 / user / mo Core · 5.00 Professional The closest paid comparison, and roughly an eighth of Haven Team per seat.
- Olvid Free · €4.99 / user / mo · €9.90 Business French, formally verified protocol, free tier that needs no identifier. Cheaper than Haven at every tier.
- Wire Free to 5 · €7.45 / user / mo annual, €8.94 monthly Mature, audited, self-hostable, with real forward secrecy. Cheaper than Haven Team.
- Proton $3.99–$29.99 / mo consumer · $12.99+ / user / mo business A whole encrypted suite — mail, calendar, drive, VPN — not just a messenger, and cheaper than Haven Business per seat.
- Element (Matrix) Community free · Enterprise quote only, from 100 seats Open federated protocol you can host yourself. Element withdrew its published per-user pricing; there is no list price to compare against any more.
- Silent Circle · Salt · Cellcrypt · KoolSpan · AWS Wickr The high-assurance end. Our competitor research did not cover these, so this page makes no claim about their prices or their designs: ask each vendor for both. What Haven does not have, whatever they offer, is forward secrecy by default, metadata minimisation and a signed native client: its forward secrecy is an opt-in per-message hash ratchet plus an opt-in, bounded deletion of superseded keys — not a Double Ratchet, and with no recovery after a key is stolen.
List prices published by each vendor, checked August 2026; they change without telling us, and currency and tier details differ. Proton renders its prices in script rather than in the page, so those figures come from a third-party tracker rather than from Proton. Check them yourself before deciding.
Where Haven sits
Above every published price in the category, and below the quote-only vendors whose guarantees it does not match. Haven is more expensive than Threema Work, Olvid, Wire and Proton, and the three products closest to its stated purpose are free.
What the money buys is operational assurance: a hosted instance somebody keeps patched, backed up and running, with named accounts, real storage, and a person who answers. That is a genuine service and it costs real money to provide — Signal publishes its own numbers, and running a messenger at scale ran about $14M a year in infrastructure alone. It is not a stronger cryptographic guarantee, and nothing on this site will tell you it is.
Use Signal, Session or SimpleX if you can. Choose Haven if you specifically need named accounts, your own rooms, larger files, and a server somebody else operates — and after reading the threat model, because the metadata Haven keeps is the limit most worth understanding. Running it yourself is still possible and still unsupported: it moves the trust rather than removing it, and makes you the operator instead of us, which is a different job and not a smaller one.
The ones worth asking
Do I need to give you an email address or a phone number?
No, and it is stronger than "not required": there is nowhere to put one. The sign-up form
has no email field and no account has a column to hold one — users.email
was dropped outright, along with every address already stored in it — and there is no
phone number field either. "We do not have your email address" is a fact about the schema
rather than a promise about our intentions, which is the version of that sentence worth
anything. The exception is a card payment: Stripe's notification about it, which we keep
verbatim, carries the email address you gave Stripe.
The privacy policy lists what else it carries.
The trade is the one that was always there: there is no password reset and there cannot be, because your private key is wrapped with your password and we have nothing to reset it against. An email address would never have changed that. What it would have done is give us one more identifier to lose or be made to hand over. Two things it does not buy you: your username is still yours and still searchable, and your connection still carries an IP address to our reverse proxy.
What can you see about who I talk to?
Who, yes; what, no. Conversation membership is a plain, indexed table on our server, and a single query joining it to itself lists every pair of accounts that talk to each other. Message timestamps are stored to the hour (the sender's signed time is to the second or finer on messages from apps not updated since 17 November 2026), and attachment sizes in bytes. We cannot read a word of what is said, but we could map who you talk to and when. If that matters more to you than the content, it is worth knowing before you choose, which is why it is answered here rather than tucked into a policy.
Does Haven have forward secrecy?
Partly, if you go and turn it on. There is an optional per-message ratchet, off by default: switched on for a conversation, every message gets its own key, your device destroys each message key 90 days after first reading or sending it and each superseded conversation key after its grace period, and while it is on only text can be sent and messages cannot be edited. It is a hash ratchet, not a Double Ratchet, so it does not heal after a key is stolen, and it does nothing about the copies the server keeps. Those are the other half: a conversation key that rotates on a schedule, and a setting — per conversation, off by default, off for every conversation that already existed — that tells the server to delete your wrapped copy of each superseded key after a grace period. Leave it off and the server keeps every historical key, wrapped to your long-term identity key, so an adversary who takes the database and later cracks your password gets everything you have ever sent rather than a window. Turn it on and the window is roughly five weeks, it depends on the conversation actually rotating, the other participant has to turn it on too or their copy still opens the thread, and history you lose is lost for good. Signal (a Double Ratchet, now alongside a post-quantum ratchet) and Wire (MLS) give forward secrecy by default, with none of these conditions attached. If that is what you need, use one of them. The threat model states every limit of ours.
Can you read my messages?
No. Message bodies, file contents, filenames and new group names arrive already encrypted, and the keys we hold are wrapped under passwords we never see. The honest caveat is the one at the top of this page: we serve you the JavaScript that does the encrypting, so a compromised or compelled server could serve a modified copy of it to one targeted person, and that person could not tell. That is a property of every browser-delivered encrypted app, ours and everyone else's, and it is the thing to weigh before trusting any of us.
What does the Free tier's 90-day limit actually do?
After 90 days, the server deletes its copy of a message. That's a cost limit, not a privacy feature dressed up as one, and it works by the same mechanism as disappearing messages — with the same caveat that applies to every disappearing-message feature in every app: it cannot reach into a device that already synced the plaintext and remove the copy sitting there. Paid plans keep the server's copy until you delete it.
What happens if I forget my password?
You lose the account. Your private key is wrapped with your password, so there is nothing for us to reset it against — a reset link would mean we held a way in, which would defeat the point. There is also nowhere for a reset to be sent: accounts hold no email address, because the column that used to hold one has been dropped from the database. Encrypted backup and restore exists, under a separate passphrase of your choosing, and lose that one and the backup is gone too.
Do you track me on this website?
No. No analytics, no cookies, no third-party scripts, no embedded fonts fetched from anyone else's server, no pixel in the footer. The only thing this site stores is your light/dark preference, in your own browser. For a privacy product, a tracker here would be self-defeating.
Can I self-host?
Yes, and it is unsupported. The source is all here and readable, and an instance you
deploy runs the same code we do — but it is not the product we sell. No support, no
managed upgrades, no SLA, and the backups, TLS, patching and disk space are yours. The
plan limits still apply, because they come from the same catalogue the code enforces, and
every account on a fresh install starts on Free — including its 90-day sweep of the
server's copy of messages. Haven serves an Onion-Location header, so an
instance you publish as a Tor onion service advertises itself and Tor Browser will move
readers to it — which is the only way to keep client IP addresses away from the server.
What changes on the trust side is worth being precise about: it doesn't remove the problem of trusting whoever serves the JavaScript, it moves that problem onto you. Hosted Haven asks you to trust an operator; running it yourself asks you to be one. Running a server well is a real skill, and for most people that swap is the worse trade rather than the better one.
Is there a group call feature?
Yes, for up to six people in a group conversation, voice or video. Six is a deliberate limit: every participant connects to every other, which keeps the call end-to-end encrypted with no media server, and costs upload bandwidth that grows with each person. That holds while the keys are genuine: each person's connection details are signed with a key the server hands out, so until you compare safety numbers your device trusts whatever key the server gives for each member and tells you when one changes. When a member you have verified changes key, calling in that group is paused, on the device where you verified them, until you acknowledge the change there; compare the safety numbers before you do. Calls have what you would expect — raised hands, reactions, speaker view, screen sharing, a waiting room for locked calls, captions and background blur — and the hands, reactions and captions travel inside the encrypted call, never through our server.
Can I pay anonymously?
Yes, with a prepaid voucher code, and that is the only way. A card ties your legal identity, your billing address and your bank to a Haven username, and that record can say more about you than any message the server holds, because the messages are encrypted and the invoice is not.
A voucher is a 24-character code that buys a set number of months on a set plan. You ask us for one, we agree how you are paying — cash, a bank transfer, Monero, a reseller who never tells us who you are — and you redeem the code in the app. Redeeming it contacts no payment processor, creates no customer record, and asks for no address. This website cannot sell you one on the spot and does not pretend to; a person mints it and sends it the way you asked.
The limit, because a claim with no limits stated is worth nothing. One column records that your account redeemed that code — the server has to know whether a code has been spent — and it is on your own data panel rather than hidden. So the unlinkability is between Haven and you, and it holds only if the code changed hands somewhere we cannot see. Buy a voucher from us by card in your own name and we can join both ends ourselves. The pricing page says all of this in full.
Will you email me about my bill?
No, and we could not if we wanted to: no account holds an email address, because the column that used to hold one was dropped from the database. There is no receipt from us, no renewal reminder, no card-expiring warning and no notice before a plan lapses. Stripe may send a receipt to the address you typed on Stripe's page — that is Stripe's message and Stripe's record. We do not send it and we cannot. We do hold that address, though, if you paid by card: Stripe's notifications are stored here exactly as they arrive so their signatures stay checkable, and a card payment's notification carries what you typed on Stripe's page. This paragraph said otherwise until an audit of the schema caught it.
The account page in the app is where the state of your subscription lives. If a card expires, nothing reaches you: the plan goes past due and free limits apply to new activity until it is settled, and nothing is deleted. Exactly what does and does not happen then is on the refunds page.