Tool calling
The agent can do things, not only say them
End a call, hand over to a colleague, press keys through another company's phone menu, or look something up in your own system while the caller waits.
Your endpoint, your data
Point a tool at an API you already have. No connector per vendor, no waiting for us to build an integration.
A small model cannot produce an invalid call
The model's choice is constrained so a malformed tool call is impossible to emit, not merely unlikely. Arguments are checked before anything reaches your system.
It costs nothing when unused
An agent with no tools makes exactly one model call per turn, as it did before tools existed. A caller hears every millisecond.
Secrets are referenced, never stored in the tool
API keys are write-only: they can be replaced but never read back, so a hijacked dashboard session cannot exfiltrate them.