What we keep,and what we don't.

Every claim below names the part of ZenLLM that proves it. Where there is an exception, it is printed next to the claim, not in a footnote somewhere else.

prompt fields in a request row

0

prompt fields in a request row

times a key is ever shown

1×

times a key is ever shown

longest a session lives

24h

longest a session lives

charged for failed attempts

$0

charged for failed attempts

One row per request. No words.

Your prompts and outputs are not written to our database or our logs. A request leaves one row of counts, sizes and prices, enough to bill you and show you a receipt.

requests1 row per call

Kept

  • Time of the request
  • Model and provider
  • Input, output and cached token counts
  • Cost, and the rates it was priced at
  • Timing and status
  • Request id
  • Size in bytes and message count
  • A fingerprint of your IP address

Not kept

  • Your prompt text
  • The model's output
  • Your IP address itself
Exceptions, in full. If a model provider returns an error, we keep that provider's error reply, encrypted, up to 8 KB, to fix routing. Dashboard Chat, if you use it, saves your conversations, encrypted. Requests still pass through model providers, and we do not promise zero retention by every provider.

Source: the request table's columns (migrations 0001, 0009, 0012, 0013, 0024, 0108), the request log line (server main.rs), provider failure records (failures.rs, migration 0032) and Chat storage (migration 0021).

API keys

A key is a secret you hold. We hold only enough to recognise it.

6

limits you can set on every key

  • A key is shown once, when you create it.

    API keys are one-time secrets (migration 0007).

  • We keep a fingerprint of the key, never the key itself.

    Key storage (migration 0001, key fingerprint helper).

  • Each key can be limited to chosen models and chosen IP addresses.

    Server-side key checks (edge apikey, request extractors).

  • Each key has its own requests-per-minute limit and an optional spend cap.

    Rate and spend checks (extractors, billing reservation).

  • Keys can expire on a date, or be frozen instantly.

    Key status checks (edge apikey).

  • These limits are enforced on our servers, not in your code.

    All of the checks above run server-side.

Your account

Signing in is guarded, and a session never outlives a day.

24h

hard limit on every session

  • Two-step sign-in with an authenticator app, plus single-use recovery codes.

    Two-step sign-in (migration 0027, edge mfa).

  • Sessions end after 24 hours, no matter what.

    Session lifetime setting (edge env).

  • Changing your password signs out every other session.

    Password change handler (edge dashboard routes).

  • Sign-in has a bot check and per-address limits.

    Sign-in guards (edge auth guards).

When a provider fails

Failures are our problem to route around, not yours to pay for.

$0

for an attempt that failed

  • If a provider is busy or fails, the request moves to the next one automatically.

    Router failover (router, mod.rs).

  • Failed attempts are never charged.

    Failure records (failures.rs, migration 0032).

  • An empty answer is retried elsewhere, not billed.

    Empty response handling (router).

  • Refusals and invalid requests come straight back, with no blind retries.

    Router error paths.

  • Every request has an id, and every error carries it.

    Request ids (migration 0009, failures.rs).

Your money

You pay for tokens a model actually produced. Nothing else.

$1

smallest top-up, credits never expire

  • Credit is reserved when a request starts and settled on the tokens actually used. The difference returns to your balance.

    Credit reservation and settlement (billing.rs).

  • Unsettled holds are released automatically.

    Reservation sweeper (billing.rs).

  • Top-ups from $1, up to $10,000. No subscription, no monthly minimum.

    Payment limits (edge payments).

  • Credits do not expire.

    The credit ledger has no expiry; Terms.

  • Refunds within 7 days: full if unused, 50% of the remainder if partly used.

    Terms of Service, refunds section.

  • Every request's cost can be rebuilt from its stored token counts and rates.

    Rate snapshots per request (migrations 0009, 0012).

What we do not hold.

ZenLLM does not hold the following certifications. They are issued only after an independent audit of our systems, and we will not display a badge for an audit we have not completed.

  • Not held: SOC 2
  • Not held: ISO 27001
  • Not held: PCI DSS
  • Not held: HIPAA

Because of that, ZenLLM is not suitable for regulated data such as protected health information or cardholder data. Those categories must not be sent through the service.

Found a problem?

Tell us on Discord what you found and how to reproduce it. Please do not test against other customers' accounts, run automated scanners that degrade the service, or access data that is not yours.