AI Keys
Promptly runs your assistant on your own AI provider account. You create an API key at OpenAI, Anthropic, Google, Zhipu, or Moonshot, paste it into Dashboard → Settings → AI Keys, and every reply the bot writes is billed by that provider directly to you. This is called BYOK — bring your own key.
There is no shared Promptly model key behind it. A workspace without a working key of its own cannot chat, during the 14-day trial exactly as much as after it. That makes this section the first stop for any new workspace, and the first place to look when the assistant stops answering.
The three-minute version
- Create an API key at your provider and make sure the account has credit or a payment method.
- Settings → AI Keys → Add key, pick the provider, paste the key.
- Leave the key field — Promptly asks your provider which models the key can use and fills the Chat model list from the answer.
- Add & verify. Watch the card: it should land on Verified with a green Tool calling ✓.
That's a working assistant. Everything below is detail you only need when something is unusual or something breaks.
Start here
- Why you bring your own key — what BYOK means for your bill, your model choice, and your message volume, and why a key is required during the trial too.
- Adding a key — the Add modal step by step, what happens automatically after you save, and how to change or rotate a key later.
- Supported providers — the five providers, their default endpoints, where to get a key, and the one-key-per-provider rule.
Configure
- Choosing a model — how the model list is fetched live from your key, what the four hints under the picker mean, what gets filtered out, and what a Model unlisted badge is telling you.
- Model roles — one key can drive three jobs (Chat, Planner, Background). Which need tool calling, and which model to point at each.
- Base URL override — for a corporate proxy or LLM gateway, plus the internal-address restrictions the override is subject to.
Keep it healthy
- What verification checks — the two real round-trips Promptly makes (connection and tool calling), the three possible outcomes, and what resets a verified key.
- Key statuses — every badge a credential can wear (Unverified, Verified, Invalid key, Out of credits, Provider down, Disabled), what caused it, and what to do. Includes the red degradation banner.
- Backup keys & failover — how the ordered key chain works, which failures trigger a failover (and which deliberately don't), and how to build a chain that actually holds.
When something's wrong
- Troubleshooting keys — the "Your assistant is inactive." banner, the visitor-facing unavailability message, and a translation table for every provider error you're likely to see.
🖼️ [Image] — The AI Keys page with a verified primary credential and one backup, showing the Tool calling ✓ badge and "Last verified" timestamp.
Two things people get wrong
- Unverified is not "probably fine." An unverified credential is skipped exactly like a broken one. After pasting a new key, verification has to pass before that key is used at all.
- Tool calling isn't optional. Every answer is planned through tool calls — knowledge search, catalog lookups, escalation. A key that connects but reports Tool calling ✗ is usable for background jobs only, and chat will ignore it. See Model roles.