Assistants

What an assistant is made of, and which parts you can influence.

An assistant is a persona, a set of capabilities, a knowledge scope and a goal. It is configured by the customer in their dashboard — you influence it only through the tools you offer.

Capabilities

Toggles on the assistant. Each one contributes prompt text and, sometimes, tools.

Capability	What it adds
Lead details	Fields the assistant must collect, one question at a time
Voice transcription	Voice notes become text before the turn
Bookings	Services, real availability, and creating an appointment
Connected tools	Your tools — see Tool servers

Capabilities are defined by the platform, not by customers. Your product does not add a capability; it adds tools inside Connected tools.

The two turn paths

Tool turn — when the assistant holds any tool

Function calling, up to six steps. The model may call several tools before it replies, and the reply itself is a tool call (finish_turn).

Chat turn — when it holds none

A single model call, no tools at all. Knowledge is retrieved once, up front, and cannot be searched again.

Enabling one of your tools moves an assistant onto the first path. That is a behavioural change, not just a new ability.

What the assistant is told about your tools

Their names, their schemas, their (sanitised) descriptions, plus three rules the platform adds:

• Call a tool rather than guessing. Never state a price, stock level, availability or order status a tool did not just return.
• If a tool errors or returns nothing, say so and offer a follow-up. Never invent the answer it failed to give.
• Take an action that changes something only after the customer clearly confirms it.

When no tools are connected

The assistant is told plainly that it cannot look up, price or order anything, and must never invent a product, price or payment link. That guard steps aside as soon as a tool server is connected — otherwise the prompt would be arguing with its own schema.