Policy before egress.
Per-key rate limits, budget caps and per-request usage logging — enforced on every request, in under a millisecond. Bad requests die here, not on your invoice.
Two guards, every request.
Every request transits the same two-guard policy pipeline before it reaches a model. Watch a live request flow through — pass or deny, in under a millisecond.
A runaway loop can't run up the bill.
Set spend caps per user, team or API key. Enforced before a token is generated — when the budget is gone, the request stops, not your card.
| User | Requests | Spend | Budget | Status |
|---|---|---|---|---|
| ARAna Ribeirogrowth | 12,481 | $142.80 |
48% of $300
|
✓ active |
| MCMarcus Chensupport | 9,032 | $96.10 |
32% of $300
|
✓ active |
| PNPriya Natarajanengineering | 22,910 | $277.20 |
92% · cap warning sent
|
⚠ near cap |
| JDJonas Dahlcontractor | 18,402 | $300.00 |
100% · hard stop
|
✕ blocked |
Per-key throttle,
token-bucket fair.
A process-wide token-bucket limiter isolates each key's throughput. A single key — or an invalid-key flood — can neither run up upstream cost nor hammer your database.
- ✓Per-key isolation. One key's burst doesn't starve another's queue.
- ✓Configurable.
requests_per_minuteper key, 0 disables. - ✓429 before cost. Throttled at the edge, before a provider is touched.
Every request, accounted for.
Every request, policy decision and tool call is logged with tokens, cost and model — visible in the dashboard, exportable to your SIEM.
Built for the auditors.
Every control your security team needs — encryption, isolation, no-training — on by default.
SOC 2 · in progress
Type II audit underway. Controls mapped to the five trust principles, evidence collected continuously.
No training on your data
We don't sell a model. We route to the best one — and never train on your prompts, ever.
Encrypted in transit
TLS 1.3 end-to-end. Credentials brokered at call time, encrypted at rest — never in prompts or logs.
Prompts not retained
Prompts pass through, get governed, and leave. We store usage metadata, never your prompts.
Put policy before egress.
Rate limits and budget caps, every request, under a millisecond. Bad requests die here — not on your invoice.