Tools
Let the assistant act during a conversation — built-in tools, your own HTTP endpoints, call transfers and long-running asynchronous work.
Tools are what the assistant can do: check a calendar, create a booking, look up an order, transfer the call. The model decides when to call one; the platform runs it and hands the result back into the conversation.
Kinds of tool
| Kind | What it is |
|---|---|
| Built-in | Provided by the platform — calendar booking, recording control, transfers and more |
| Integration | Provided by a connected integration, using its credentials |
| Custom | Your own HTTP endpoint, described with parameters the model fills in |
A built-in tool can also be copied into a flow and adjusted there, without changing it for every other flow.
Custom HTTP tools
A custom tool is an HTTP request with a name, a description the model reads to decide when to use it, and parameters it fills from the conversation. The response comes back to the model, which uses it in its next reply. See Custom HTTP tools.
Custom tools can be managed over the API — see Custom tools — and tested against the echo endpoints before you point them at a real system.
Tip
Write the tool description for the model, not for a developer: when to use it, what each parameter means, and what a good answer looks like. It is the most effective lever on whether a tool gets called at the right moment.
Long-running work
Some actions take longer than a caller should wait in silence. An asynchronous tool returns straight away and reports back later, through a callback. What happens when the result arrives is configurable:
- acknowledge only, or just confirm the action ran;
- say the result into the conversation;
- add it to the model's context, for it to use when relevant;
- add it silently, without the assistant mentioning it.
Callbacks authenticate with scoped credentials, so an endpoint that reports back can do only what it was given permission for. See Asynchronous tools.
Safety on voice
Voice transcripts can be wrong, and some actions should not run on a half-heard sentence. A tool can require a confirmed transcript before it executes, and the assistant can say a short preamble before calling it, so the caller knows what is about to happen.
Reusing logic
Scripts shared by several tools and automations can be kept as global scripts and referenced from each, instead of copied into every one.