Skip to content

Rate limits

Limits count per environment unless they say otherwise, in fixed windows of the length given. A limit lets its count through in each window, so across the end of one window up to twice the count can pass in a short time.

What is counted Per Limit Over it
Login steps Caller’s IP address 30 in 10 minutes 429
Login steps on one account Account 10 in 10 minutes A captcha for every caller, when the platform has a captcha; otherwise 429. Authenticator-app and recovery codes are always refused (429) past 10; a device the account has signed in on before has 10 of its own, which nobody else’s tries spend.
Answers to an authenticator-app or recovery code Account 20 a day from new devices on new networks Every answer counts before it is checked, right or wrong. Whoever answers has already passed the first factor. Once the 20 are spent, the networks the account has a session from share 10 more, and the devices the account has signed in on share 10 of their own (then they go on as any device), so its owner still gets in at home, at work or on a device they used before. Past those, the next 20 answers are checked but never accepted: a wrong one is invalid_code, and the right one is answered second_factor_paused (“two-step codes for this account are paused for up to a day”). A right authenticator code is spent by that, a recovery code is not. After that no answer is checked. The account’s owner gets one email at the fifth wrong answer of a day. A passkey as the second step is not held up by any of this.
Tries of a secret without a captcha that did not succeed Caller’s IP address, and account 3 in 10 minutes A captcha for every step that tries a secret. Each try is counted before it is checked, so tries sent at once cannot all slip under it.
Answers to email codes Address 20 a day from new devices on new networks Every answer counts before it is checked, right or wrong. Once the 20 are spent, the networks the account has a session from share 10 more, and the devices the account has used share 10 of their own (then they go on as any device), so its owner still logs in with a code at home, at work or on a device they used before. Past those, the next 20 answers are checked but never accepted: a wrong one gets the answer every wrong code gets (invalid_code), and the right one ends its code and is answered codes_paused (“codes for this address are paused for up to a day”), which only the person holding the code can see. After that no answer is checked. Only answers to codes that were really sent count, and nothing in the answers says whether the address has an account. A password, a passkey or a provider still logs in. The account API’s change of address counts its answers here too, on the address it moves to.
New login attempts (each GET /authorize that shows the login page, each new embedded login) Caller’s IP address 200 in 10 minutes 429
New passwords (a sign-up’s, a reset’s) Caller’s IP address 10 in 10 minutes, and 30 across every environment 429
Passwords checked without a captcha Caller’s IP address, across every environment 100 in 10 minutes A captcha for the next password logins from that address, never a refusal. Without a captcha on the platform, only the login steps’ limit of each environment applies.
New wallet accounts Caller’s IP address 10 an hour 429

A login step that tries a secret is a password, an email code, an authenticator-app or recovery code, or a wallet or passkey signature.

What is counted Per Limit Over it
Codes sent Caller’s IP address 10 in 15 minutes 429
Codes asked for without a captcha Caller’s IP address, across every environment 30 in 15 minutes, and 100 for a whole IPv6 /48 A captcha for the next codes asked from there, never a refusal: everyone behind one address (a carrier, an office, a VPN) shares it. Without a captcha on the platform, only the limit of each environment applies. A request that reaches SeekAuth without the address its proxies report is counted in its environment only.
Codes asked for one mailbox without a captcha The mailbox, however its address is written: with or without a +tag at any provider, and with dots at Gmail 3 in 15 minutes A captcha for the next codes. Counted before the code is sent, so requests sent at once cannot all slip under it.
Codes asked for one address Inbox 10 in 15 minutes (3 without a captcha on the platform) 429. A code that is sent and one that is only answered for because of the address’s state (no account and sign-up closed, a reset for nobody) count alike here. A re-authentication that names an address other than its user’s, and the account API’s change of address, count only against their caller.
Codes asked for one address from a device the account has used Address and device 3 in 15 minutes 429. This budget takes the place of the two limits above, so a known device is not held up by somebody else’s attempts.
Codes sent to one inbox The inbox, across every environment 10 in 15 minutes, for codes to an address the environment has no account for (sign-up codes, an address being added or moved to), to an account whose address SeekAuth has not proven, and “an account exists” notices to such an account The answer is the same and the code is held back. Asking again (Resend) in a later window sends it, unless that window’s 10 were spent first. Each held-back code is logged with the environment and a hash of the inbox, never the address. Mail to an account whose address SeekAuth has proven is never held back here: only the limits of each environment apply to it. An address is proven once its owner answered a code sent to it (a login, sign-up, verify or reset code, or a change to it through the account API), or linked Google or Apple with it verified. Until then it is not: an account a custom provider made, an account whose address was never verified, and a Google or Apple account whose address was changed through the account API before November 2026.

| Codes asked for the +tag spellings of one mailbox without a captcha | The mailbox behind the tag (name+anything@ counts for name@), across every environment | 10 in 15 minutes | A captcha for the next codes to any of its tag spellings; nothing is held back. Counted apart from the address without a tag: codes asked for tag spellings take nothing from that address’s limits across environments. | | Codes asked for one domain without a captcha | The domain, across every environment, unless it is a large public mail provider’s (Gmail, Outlook, iCloud and the like) | 100 in 15 minutes | A captcha for the next codes to any address there, whoever has it; nothing is held back. A domain can be one mailbox that takes every address, which no limit by address would bound. |

An inbox is counted by the address it reaches, so the ways of writing one Gmail address count together. The limits across environments exist because environments are free to create.

What SeekAuth sends in 15 minutes without a captcha is therefore bounded: from one caller, 10 codes in one environment and 30 in all of them; to the +tag spellings of one mailbox, 10 from all callers together; to one domain that is not a public provider’s, 100 from all callers together. Past those, each code costs whoever asks for it a solved captcha, and a caller still gets at most 10 in each environment. To one address with no proven account, 10 from all callers together are sent and the rest held back (above). None of the counts by a domain, a mailbox’s tag spellings or a network holds back or refuses anybody’s code: a person who shares them with whoever spent them solves a captcha and gets their code.

An inbox is counted across environments only by mail that is really sent. Where no code goes out, the answers are the same as for one that does, and the step writes as much as one that sends a code, so that neither what is answered, nor where a refusal falls, nor how long a step takes says whether the address has an account.

The account API counts per session, never per address: a web client’s backend sends every user’s calls from one address, and somebody else signed in to the account (whoever holds a session of it) must not be able to spend what its owner needs to put them out. What only takes access away is not counted at all.

A call that needs a fresh login is counted only once it has shown one, so a session that could not make the change cannot use up the limit of the person who can.

Route Limit
Changes that add to what guards the account: POST /account/v1/password, POST /account/v1/totp/confirm, POST /account/v1/recovery-codes, POST /account/v1/passkeys, POST /account/v1/identities/link and …/link/confirm, and deleting a developer account 30 an hour per session, all of them together
What only takes access away: DELETE /account/v1/password, DELETE /account/v1/totp, DELETE /account/v1/passkeys/:id, DELETE /account/v1/identities/:id, DELETE /account/v1/sessions/:id Not counted. Ending another session than the token’s needs a fresh login.
POST /account/v1/password Also 10 in 10 minutes per session, counted before the new password is screened and hashed. A password refused for its length is not counted.
Other writes: PATCH /account/v1/profile, PATCH /account/v1/passkeys/:id 120 an hour per user
POST /account/v1/email (send a code to a new address) 10 in 15 minutes per session, taken address or not, and 3 of them to one domain that is not a public provider’s or to the +tag spellings of one mailbox; nothing is counted on the address in the environment. A code to a free address counts as mail to a new address across environments (10 per inbox in 15 minutes), past which the answer is the same and no code is sent (ask again once the window is over). An address that is taken gets nothing.
POST /account/v1/email/confirm 10 in 10 minutes per session. Each code also allows 5 tries, and its answers count on the address’s answers to email codes for the day (above), so a code sent to somebody else’s address cannot be guessed at length.
What is counted Per Limit
Creating workspaces Developer 30 an hour
Creating apps Workspace 30 an hour
Sending invitations Inviting developer 30 an hour
Invitation emails to one address The mailbox the address reaches (with or without a +tag, and with dots at Gmail), across workspaces 5 a day; past it the invitation is still made and shown when the person signs in, without an email
Moving a group, or an app on its own, to another slug Workspace Counts as creating an app: the address it leaves stays held for the workspace
Checking a payment (POST …/billing/confirm) Workspace 30 an hour
Starting a checkout (POST …/billing/checkout) Workspace, and the developer across workspaces 10 an hour per workspace and 20 per developer. A checkout that is still open is given again for 15 minutes and counts nothing.
Cancelling (POST …/billing/cancel) Workspace, and the developer across workspaces 10 an hour per workspace and 20 per developer

A developer may also own at most 10 workspaces (409 past that).

An API key that does not exist is counted nowhere. The keys that exist are known to each server, so a made-up key (or a string that is not shaped like one: sk_dev_ or sk_live_ and 43 characters) is 401 at once, costs nothing, and cannot make anybody’s valid key wait, whatever address it comes from. A key made a moment ago works at once on the server that made it and within five seconds on any other. A key that was revoked is 401.

The status is 429 with a Retry-After header, in seconds.

On the OAuth endpoints and the embedded login, the body is an OAuth error:

HTTP/1.1 429 Too Many Requests
Retry-After: 540
Cache-Control: no-store
{ "error": "temporarily_unavailable", "error_description": "too many attempts, try again later" }

On the account API and the management API, it is an API error:

{ "code": "rate_limited", "message": "Too many attempts, try again later" }

A captcha asked for instead of a refusal is not an error: the login answers with its challenge as usual and the rejection code captcha_required (see Rejections). The SDK’s login UI shows the captcha and sends the step again with it.

  • An IPv6 caller counts by its /64 network; the captcha on codes across environments also counts its /48.
  • Behind proxies, the caller’s address is read from the proxies the deployment trusts, never from what a client writes itself.
  • Nobody can keep a known user out by spending a shared limit for free: with a captcha on the platform, an account tried too often, or an address sent its 3 codes, asks every caller for a captcha instead of refusing; a device that has held a session of the account has codes of its own; a code that is never sent cannot be used to spend an address’s codes, its ceiling in an environment, or its answers; filling an address’s ceiling in an environment takes real codes in its owner’s inbox there; and no answer says whether the address has an account.

Checking a password is deliberately slow. At most eight run at once per server; a check that has to wait more than a couple of seconds is refused with 429 and Retry-After: 1, whether or not the address has an account.

A caller may have 100 passwords checked in 10 minutes across every environment without a captcha, and is then asked for one; in one environment, 30 login steps (above).

A new password is screened against breached passwords and hashed only where the step can go on: never where the environment has no password login, a sign-up’s only once its captcha is solved (when the platform has one), and a reset’s only once its code has been proven. It is counted as a new password before any of that.

Notices (a new device, a changed password, a changed second step and the like), invitations and welcomes are sent apart from codes and never ahead of one. While a notice waits to be sent, another one that says the same to the same address is not sent as a second email: the one that goes out says how often it happened. An address has at most 20 notices waiting in an environment. A code that would be sent after its ten minutes have passed is not sent.