FAQ

Check three things:

1. The key starts with nk_.
2. You are passing it as the api_key parameter (OpenAI SDK) or x-api-key header (Anthropic SDK).
3. The key is still active — go to API Keys and confirm it hasn't been revoked.

No — only the SHA-256 hash is stored. Revoke the lost key on the API Keys page and generate a new one. No calls can be made with a revoked key.

No. Once an admin has connected at least one provider key, any team member can create their own Zelyx key from the API Keys page. You only need admin help if no provider key has been added yet.

Yes. Create as many as you need — one per environment, machine, or project. Each key is attributed to the user who created it, so per-person analytics still work correctly.

Budget enforcement is multi-layered. A call can be blocked by any of the seven layers even if others have headroom. Check:

• Per-model budget — the specific model may be paused or capped.
• Per-key budget — the Zelyx key may have a daily budget set.
• Team or project budget — if the call is tagged, that layer may be exhausted.
• Run budget — if using X-Zelyx-Run-Id, that run may be exhausted.

The Explorer page shows the gate decision and reason for any blocked call.

No. When Zelyx blocks a call (HTTP 402 or 403), the request is never forwarded to the provider. No tokens are consumed and no provider cost is incurred.

The exception is Review required (HTTP 200 + X-Zelyx-Review-Required: true) — those calls do reach the provider and incur cost, because the call proceeds.

Zelyx charges a platform fee (see the pricing page). Your provider costs are charged directly by the provider to your API key — Zelyx does not mark up or resell tokens. You bring your own key; you pay the provider directly.

Yes, a small amount — typically 5–20ms for key lookup, gate evaluation, and forwarding overhead. Provider latency (often 200ms–2s+) dominates in practice. Check the Latency page for P50/P95 figures on your actual traffic.

Yes, fully. Pass stream=True exactly as you would with the provider SDK. Zelyx always streams from the provider internally (for token and TTFB measurement), and adapts the response format to what your client requested.

The call completed and you received the AI response. The flag is informational: it means the gate detected something worth a human look — typically a high-cost model, a payment-related prompt, or a suspicious tool call.

Review the call on the Gate page. If it's legitimate, mark it approved. If you see this too often for normal traffic, check your gate configuration or contact support.

Use a unique session ID per batch run (e.g. batch-2026-06-04-a3f2) so batch calls are isolated from interactive sessions. You can also increase the loop threshold in Settings if your batch volume is predictably high.

If the action is set to warn, the alert fires but calls proceed — no disruption. If set to block, your batch will hit 429s; switch to warn while tuning.