ai_profiles

What the backend will accept: which providers this deployment can run and their voices, every enabled model per kind, and each configuration class with the fields and bounds a component of it carries. Use before building a profile or a component, to know what values are valid. Not organization-scoped — actor is required only so every route on this surface carries the same authentication, not to select what comes back.

List the Live (speech-to-speech) components this organization can use: its own and the shared templates, each with the count of profiles pointing at it.

Create a Live component: one speech-to-speech model that hears, decides the turn, thinks and speaks by itself -- an enabled live model, one of its provider's live voices, the language it listens and speaks in, and its thinker where its class offers one. Use when a profile's voice section should run as one live model instead of separate STT/LLM/TTS/VAD components. It names no standby: its fallback is the next arrangement of a profile's voice section.

Delete one of this organization's Live components. Refused with the count while any profile points at it.

Edit a Live component; only the fields sent change. If more than one profile uses it, send apply_to=all to change it for all of them, or apply_to=copy with profile_id to change a copy for that profile only.

List the language model components this organization can use: its own and the shared templates. Each carries in_use_by (profiles pointing at it) and standby_for (components naming it as their standby).

Create a language model component: a provider, an enabled model and the settings its class carries. Use when a profile needs a new LLM ingredient instead of reusing an existing one. max_tool_rounds is how many rounds of tools one turn may chain before it is cut; omitted, it starts at the bounded default. null means no ceiling -- a turn chains rounds until the model answers and is never cut for looping. Write that only on purpose: an agent that runs twenty tools without a word to the person is usually looping, so a client agent should keep a bound.

Delete one of this organization's language model components. Refused with the count while any profile points at it or any component names it as its standby.

Edit a language model component; only the fields sent change. Use when an existing LLM component's settings need to change. Sending max_tool_rounds: null lifts the ceiling on rounds of tools per turn (see the create operation for what that means and why a client agent should keep a bound); leaving it out keeps the ceiling as it is.

List the AI profiles this organization can use: its own and the shared templates, each with its voice arrangements and text section resolved inline and in_use_by, the number of agents with a version on it.

Create an AI profile — the bundle of model settings an agent pins. Use when an agent needs a new combination of voice/text components instead of reusing an existing profile.

Read one AI profile, own or shared template, with its voice arrangements (in the order a call tries them), its text section and in_use_by, the number of agents with a version on it. Another organization's profile is not found.

Delete an AI profile. Use when the profile should no longer exist. Refused while an agent still uses it; the refusal says how many.

Change which settings an AI profile points at. Use when an existing profile's name or pinned components need to change.

Duplicate an AI profile as a new one. Use when starting a new profile from an existing one instead of from scratch — a route of its own, never a side effect of saving an edit.

Write a profile's text section: the model every written conversation is held by, the recognizer that reads a voice note -- or none, which announces voice notes instead of transcribing them -- and the language model that reads images and documents, with its ceilings. A reader or ceilings left out are kept as they are; reader_id: null removes the reader. Use when setting up or changing how this profile handles written conversations.

Write a profile's voice section whole: its arrangements, in the order a call tries them. Use when setting up or changing how this profile handles calls, including its fallback order. The first that can open takes the call; each is either a pipeline (recognizer, model, voice, detector) or a speech_to_speech arrangement (one Live component). An empty list is a profile that takes no calls.

List the speech-to-text components this organization can use: its own and the shared templates, each with in_use_by and standby_for counts.

Create a speech-to-text component: a provider, an enabled model, the language it listens in and how it decides the caller has finished speaking. Values are checked against GET /ai-profiles/capabilities; standby_id names the component a text profile's transcriber falls back to -- on a call the standby is the recognizer in its place in the voice profile's next arrangement below.

Delete one of this organization's speech-to-text components. Refused with the count while any profile points at it or any component names it as its standby.

Edit a speech-to-text component; only the fields sent change. If more than one profile uses it, send apply_to=all to change it for all of them, or apply_to=copy with profile_id to change a copy for that profile only.

List the text-to-speech components this organization can use: its own and the shared templates, each with in_use_by and standby_for counts.

Create a text-to-speech component: a provider, an enabled model and one of its voices, with language, speed, volume, emotion and an optional pronunciation dictionary. A pronunciation_dict_id the provider says does not exist is refused, and so is a standby_id: a voice's standby on a call is the voice in its place in the profile's next arrangement below, so set it there.

Delete one of this organization's text-to-speech components. Refused with the count while any profile points at it; a voice naming it as a standby no longer counts, and that standby is cleared.

Edit a text-to-speech component; only the fields sent change. If more than one profile uses it, send apply_to=all to change it for all of them, or apply_to=copy with profile_id to change a copy for that profile only. A standby_id other than null is refused: add a row below in the profile instead.

List the voice activity detection components this organization can use: its own and the shared templates, each with in_use_by and standby_for counts.

Create a voice activity detection component: how loud and how long sound must be to count as speech, how much silence ends it, and whether the caller may interrupt the agent. Every field has a default. A standby_id is refused: a detector's standby on a call is the detector in its place in the profile's next arrangement below, so set it there.

Delete one of this organization's voice activity detection components. Refused with the count while any profile points at it; a detector naming it as a standby no longer counts, and that standby is cleared.

Edit a voice activity detection component; only the fields sent change. If more than one profile uses it, send apply_to=all to change it for all of them, or apply_to=copy with profile_id to change a copy for that profile only. A standby_id other than null is refused: add a row below in the profile instead.