Create a connection (provisions inline)

A human-readable name for the connection (1–120 characters).

How the line is handled. 'hosted' (default): an in-network voice assistant (cascade STT → LLM → TTS) is the brain. 'manual': your own LLM is the brain — Saperly forwards text turns and executes the directives you return. Audio stays in-network in both modes.

The hosted brain backend. Today only 'network' (an in-network voice assistant) is available; 'openai_realtime' is disabled and rejected on create/update.

The system prompt that defines the assistant's behavior. Put any opening line here — there is no separate greeting field.

The spoken language as a BCP-47 tag (e.g. "en", "es", "en-US"). On hosted lines it is the in-network transcription hint — which language the caller speaks (pair it with a voice that speaks the same language, and instruct the model to reply in it). On manual lines it is the Conversation Relay language, steering in-network transcription AND the spoken voice. Discover the supported set from GET /languages (or filter GET /voices by language); any well-formed tag is accepted. Send "" to clear it (auto-detect).

When true (the default), the do-not-message tool is exposed and disclosure is spoken as the line's first, uninterruptible utterance (the TCPA notice). When on with no disclosure, a standard org-named default is filled in.

Hosted-line opt-in to auto-answer inbound SMS with the model. Defaults off. (Manual lines forward to their webhook brain regardless.)

The TCPA notice spoken first, uninterruptibly, when complianceEnabled is on. Leave empty with compliance on and Saperly fills a standard org-named default.

Manual-mode webhook-brain URL: where Saperly POSTs each turn (signed with the connection's manualSecret) when no live agent socket is connected. Only meaningful for manual mode; send "" to clear it.