Skip to content

Security

We hold the one password your agent can’t. Here is exactly how.

Your agent’s maker decided it should never hold a raw password, and they were right. That makes us the ones you have to trust, so this page says plainly what we do, what we keep, and what happens when something goes wrong.

The path of one call

The agent never talks to the vault

Each call is checked against your token and your rules before it goes anywhere. Only one program ever receives your password, and your agent can’t reach it.

  1. Your agent

    Holds one agent token. Asks for an action, like mail.search.

  2. Front door

    Checks the token against its hash, then your rules: is this tool switched on for this agent? The call is written to the log.

  3. Connector

    Does the work, keeps to your send limits and hides sign-in codes. It holds a handle to your password, never the password.

  4. Mail session

    The one program the vault hands a password to, so it can sign in to your provider. It has no keys and no database.

  5. Vault

    Keeps passwords encrypted and has no public address. It, not the caller, decides which server a password may go to.

Only the mail session can ask the vault for a password, and only to sign in to your provider. Nothing can show one to a browser or an agent: a stored password can be replaced or deleted from your dashboard, never read back, not even by you.

What we promise

Six things that hold on day one

These are the bar for launch, not a plan for later.

Passwords go in and never come out

A stored password can be replaced or deleted, never shown again. Automated tests plant fake secrets and fail the build if one appears in a reply, a log or an error.

A key that is yours alone

Your browser encrypts the password for the vault before it is sent, so our web servers only carry it sealed. In the vault it sits under a key made for your account, itself encrypted under a master key.

Read first, write by choice

Read tools are on from the start. Writing is off until you turn it on, and each write tool then has a switch of its own. Turn writing on with none of them chosen and they all come on, ready to be turned off one at a time.

Codes and reset links are hidden

One-time codes, password-reset links and magic sign-in links are taken out before your agent reads a message. It works by reading the text, so it catches the wordings we know and cannot promise every one. You can turn it off for a connector.

Sending has a ceiling

A daily cap, a recipient cap, a pause after a bounce and an optional list of who may be written to. A confused agent can’t spam from your address.

Off means off

Revoking a token stops calls within 5 seconds. Deleting your account deletes your passwords and tokens right away, and the audit log within 24 hours.

Your data

What we keep, and what only passes through

Nothing here is used for training or advertising, and there are no third-party trackers on this site or in the dashboard.

WhatKept?How, and for how long
Your app-specific passwordStored, encryptedUntil you delete it here. Cancel it at Apple and our copy stops working at once.
Your name, email address and passkeysStoredTo sign you in. A passkey is a public key, and there is no account password to steal.
The audit logStoredTool, time, which token, where it went, result and how long it took. 30 days on Free, 1 year on Personal.
Your switches, limits and recipient listStoredSo your rules hold on every call. Deleted with your account.
New-mail notices waiting for your agentStoredWho a new message is from, its first three recipients and its subject, with codes and sign-in links hidden while hiding is on, and never its text. Kept until your agent picks it up, for 7 days and 200 notices per mailbox at most, and deleted when you switch new-mail events off or disconnect the mailbox. Cloudflare encrypts it where it is stored; the vault, which guards your password, does not.
Webhook deliveries waiting for your agentStoredWhat a service posted to your inbox: its body, its query, and the few headers that say what it is, never a credential or a signature. Kept until your agent picks it up, for 7 days and 200 deliveries per inbox at most, and deleted with the inbox. The address holds a secret we keep only a hash of. Cloudflare encrypts it where it is stored; the vault does not.
Your inboxes’ signing secretsStored, encryptedIn the vault, as a password is, and used there only to check a sender’s signature. Never sent anywhere or shown again. Deleted when you stop checking signatures or delete the inbox.
Messages, attachments, events and contactsNever storedPassed to your agent in memory and gone when the call ends.
What your agent asked forNever storedSearch words, recipients and other inputs are never written to the log. Each call keeps a fingerprint of them, which tells two identical calls from two different ones. Nothing can be read back out of it, though somebody holding the database could use it to check a guess.
Your agent tokenNever storedWe keep a hash and the first few characters, enough to recognise it.
  • Your app-specific passwordStored, encrypted

    Until you delete it here. Cancel it at Apple and our copy stops working at once.

  • Your name, email address and passkeysStored

    To sign you in. A passkey is a public key, and there is no account password to steal.

  • The audit logStored

    Tool, time, which token, where it went, result and how long it took. 30 days on Free, 1 year on Personal.

  • Your switches, limits and recipient listStored

    So your rules hold on every call. Deleted with your account.

  • New-mail notices waiting for your agentStored

    Who a new message is from, its first three recipients and its subject, with codes and sign-in links hidden while hiding is on, and never its text. Kept until your agent picks it up, for 7 days and 200 notices per mailbox at most, and deleted when you switch new-mail events off or disconnect the mailbox. Cloudflare encrypts it where it is stored; the vault, which guards your password, does not.

  • Webhook deliveries waiting for your agentStored

    What a service posted to your inbox: its body, its query, and the few headers that say what it is, never a credential or a signature. Kept until your agent picks it up, for 7 days and 200 deliveries per inbox at most, and deleted with the inbox. The address holds a secret we keep only a hash of. Cloudflare encrypts it where it is stored; the vault does not.

  • Your inboxes’ signing secretsStored, encrypted

    In the vault, as a password is, and used there only to check a sender’s signature. Never sent anywhere or shown again. Deleted when you stop checking signatures or delete the inbox.

  • Messages, attachments, events and contactsNever stored

    Passed to your agent in memory and gone when the call ends.

  • What your agent asked forNever stored

    Search words, recipients and other inputs are never written to the log. Each call keeps a fingerprint of them, which tells two identical calls from two different ones. Nothing can be read back out of it, though somebody holding the database could use it to check a guess.

  • Your agent tokenNever stored

    We keep a hash and the first few characters, enough to recognise it.

One thing to know: whatever your agent asks for is sent to your agent, and from there it is in the hands of whoever runs that agent, under their terms. We are the carrier for that handover.

When things go wrong

What each failure can and can’t reach

Spotting that is your agent’s job. Limiting the damage is ours: read only by default, send caps and recipient lists hold no matter what the email says, because we check them and nothing your agent reads can talk us out of them. Hiding sign-in codes is not of that kind: it reads the text, so it lowers the odds a good deal and is not a guarantee.

Check our work

What you can verify today, and what is not there yet.

Every call
In your audit log, with export
Our access
Listed at Apple. Cancel it there
Dependencies
Pinned, never “latest”
Independent audit
None yet

Found a problem?

Tell us first. We read every report and reply as soon as we can.

First reply
As soon as we can
Bounty
Planned once there is revenue