Troubleshooting keys
Start on Settings → AI Keys: the card's status badge and its Last error line carry the provider's own message, which is almost always the fastest diagnosis.
"Your assistant is inactive."
An amber banner across the top of the dashboard, with an Add your AI key button.
Your workspace is on the free trial and has no usable key — meaning no credential that is both Verified and Tool calling ✓. There is no shared Promptly key to fall back on, so the assistant genuinely can't answer. The API refuses chat requests with 402 TRIAL_REQUIRES_OWN_KEY.
Fix it by adding a key (Adding a key) or by verifying the one you have. Note what does not count:
- an Unverified key (added but never checked, or reset by a re-key),
- a key with Tool calling ✗ — background-only, and chat needs tools (Model roles).
The banner re-checks as you move around the dashboard, so it clears on its own once a key verifies. Dismissing it hides it for the session; it doesn't change anything server-side.
The widget says the assistant is temporarily unavailable
The visitor-facing message when every credential in the chain has failed:
The assistant is temporarily unavailable. Please try again in a few minutes or contact support.
Open the AI Keys page — the red LLM key issue detected banner names which keys are unhealthy and why. See Key statuses and Backup keys & failover.
Provider errors, translated
| What you see | Usual cause | Fix |
|---|---|---|
| Invalid key (401/403) | Key revoked, truncated on paste, or belongs to a different provider than selected. | Re-copy from the provider console; check the provider dropdown matches. |
| Out of credits (402, or insufficient_quota / credit balance) | No payment method or exhausted prepaid balance. | Top up at the provider, then Verify key. |
| Provider down (5xx / timeout) | Provider incident, or an unreachable base URL. | Re-verify later; a verified backup key covers the gap. |
| Verified but Tool calling ✗ | This model can't call functions. | Pick a different chat model in Edit models, then Test & save. |
| Verification fails with a model-not-found error | Model name isn't available on this key — often a delisted or region-restricted model. | Edit models → Refresh, pick a listed model. |
| Showing defaults — couldn't fetch models | The model listing call failed. | Check the key and base URL; don't save on top of a failed fetch (Choosing a model). |
| Model unlisted badge | A saved model is no longer offered by the provider. | Edit models, pick a current one. |
| Adding a key is refused as a duplicate | One credential per provider. | Edit the existing credential instead of adding a second. |
🖼️ [Image] — A credential card in the "Invalid key" state with the provider's raw error text under "Last error".
Rate limits (429)
A 429 from your provider is not treated as an unhealthy key, so it doesn't fail over and doesn't change the status — the request simply fails. If visitors hit this during busy periods, raise your rate limit at the provider or move Chat to a model with more headroom. The exception is a 429 whose body mentions quota exhaustion, which is classified as Out of credits.
Still stuck
Check whether the model works outside Promptly with the same key (a plain curl to the provider's chat-completions endpoint). If it fails there too, it's a provider-side problem. If it works there but not here, note the Last error text and contact support.