AI agents with your own model key
Run your chat, WhatsApp and voice agents on your own key from the AI provider you already use (or any compatible endpoint), while our rules decide what an agent may do.
Steps
- Settings → AI keys: add a key. It is checked once, stored encrypted and never shown again (only its last four characters).
- Choose per agent which model to use, and what happens if your key fails: answer on our model, or do not answer.
- In the agent's Model and rules tab, choose its tools and limits. Every action the model proposes is checked against rules it cannot change.
- Run a rehearsal: Safety (seven scripted difficult visitors) or Business tasks (a question your knowledge should answer, an order the agent cannot see, a call-back request, a tool you withdrew, and a provider failure). The report quotes each reply, says why a case failed, and shows the measured cost per task.
- Open any conversation to read "Why the agent did this": the tools it ran, what the rules refused, handoffs and fallbacks.
Questions
- Can a visitor talk the agent into doing something?
- Messages that look like instructions are treated as data, nothing is written on such a turn, and card numbers are removed before the model sees them.
- Does a passed rehearsal mean live conversations will go the same way?
- No. A rehearsal is a sandbox: scripted visitors, nothing sent, saved or handed off. The provider failure is simulated without calling a model. The checks read what the rules and the trace recorded, not a second model's opinion, and the model's own conduct is shown apart from the outcome: a case can pass because the rules refused what the model asked for. The cost is what the run measured; on your own key your provider bills it. Live conversations are in the agent inbox, each with its own trace.
Updated 2026-11-03 · llms-full.txt