PATCHhttps://core-api.heysadie.ai/assistants/{id}

API key in the ApiKey scheme: ApiKey YOUR_API_KEY. Tenant keys authorize their own tenant; organization keys can act on any tenant in the organization.

Required when authenticating with an organization key: it selects the tenant the request acts on. Omit with tenant keys.

The resource's UUID.

Display name of the assistant (e.g., Sadie, Alex)

The voice the assistant will use to speak to callers

The transcriber engine used to convert speech to text

The LLM model that powers the assistant's reasoning (e.g., GPT-4)

Custom prompt section developers can inject into the default assistant prompt

JSON object with dynamic values injected into developerPrompt (e.g., {{opening_hours}})

Whether call recordings are enabled for this assistant

The phone number to transfer calls to if the assistant fails to respond (E.164 format)

Turn on or off the failover mode (in case it's enabled, the call will be forwarded to the failover number)

The phone number this assistant is assigned to

The first message the assistant says when answering the call

List of tool IDs the assistant can use

Timezone of the business the assistant is handling calls for (e.g., Australia/Sydney)

Address of the business (used in prompt reasoning)

Final message before the call ends

Whether the assistant can read and reference the caller's phone number if not anonymous

Prompt used to control how the assistant summarizes the call transcript

Prompt that helps classify calls based on the categories field

List of categories the assistant can classify the call into (e.g., booking, edit)

The ID of the prompt associated with the assistant

Optional chat/messaging prompt ID (type 'chat'); null clears it

Optional whisper-summary prompt ID (type 'whisper_summary'); null clears it

Optional voice persona prompt ID (type 'voice_persona', realtime models); null = runtime default persona

Whether the assistant has a background sound

The webhook URL override for the assistant; null clears it

Array of notification webhook URLs for the assistant

Whether CSAT (Customer Satisfaction) feedback is enabled for this assistant

Array of messages to send when the assistant detects the caller is idle

Transfer mode behavior for escalations. Accepts exactly 'cold-transfer', 'smart-summary', or 'warm-return'; any other value is rejected with a 400. Matched exactly and case-sensitively. Defaults to 'cold-transfer' if not provided. 'smart-summary' briefs the destination then connects the caller; 'warm-return' briefs the destination and returns the caller to the assistant if nobody answers.

Maximum call duration in seconds. Must be between 10 and 43200 (12 hours). Null means no limit.

Number of seconds to wait before considering the caller idle. Min 0 - Max 60.

Maximum number of times idle messages can be spoken before ending the call. Min 0 - Max 10.

Per-assistant settings container. Currently holds conversationConfig; additional setting categories will be added here in the future. Structure: { conversationConfig: { endpointingPlan?: { mode?: "fixed"|"dynamic", minDelay?, maxDelay? }, turnDetectionPlan?: { timeout?, unlikelyThreshold? }, vadPlan?: { minSpeechDuration?, minSilenceDuration?, prefixPaddingDuration? }, interruptionPlan?: { minDuration? (0-5), minWords? (0-10) }, deliveryMode?: "none"|"emotion_tagging"|"expressive", expressiveOptions?: { disfluencies?: boolean, nonverbalSounds?: boolean | { laughing?, breathing?, sighing?, crying?, vocalizing?, mouthSounds?, reflexSounds? }, pace?: "slow"|"normal"|"fast", instructionsAppend?: string (max 2000) } | null, emotionTaggingEnabled?: boolean (legacy; superseded by deliveryMode) } }

Dial timeout in seconds for transfer attempts. Integer between 1 and 60. Merged into settings.transferDialTimeout; other settings keys are preserved.

Pause in seconds before the warm-transfer whisper plays to the supervisor. Integer between 0 and 30. Defaults to 0 when omitted. Merged into settings.warmWhisperDelaySeconds; other settings keys are preserved.

What the assistant says when a warm-return transfer goes unanswered and the caller comes back. Max 2000 characters. Merged into settings.warmReturnMessage; other settings keys are preserved. Send "" to restore the built-in wording. Set this to match whatever your server does when it receives supervisor_answered: false on the end-of-call report.

When true, CoreProxy leaves {tool_instructions} empty and sends linked taskgroups/tasks as structured data to AgentSpawn. When false (default), the existing prompt pipeline runs unchanged.

Transfer knobs for this assistant, stored in the transfer_settings column. Structure: { amd_detection?: boolean, use_assistant_number?: boolean }. Unknown keys are rejected. This DEEP-MERGES on PATCH, so sending one knob leaves the others as stored; send null to clear the whole document. amd_detection controls answering-machine detection on a warm-return transfer destination and defaults to on when absent. Turning it off does not merely skip detection: the caller stays bridged to whatever answered, voicemail included, instead of being returned to the assistant. use_assistant_number controls the caller ID shown to the person receiving a cold or warm transfer and defaults to false when absent. When false, they see the customer's phone number, or the assistant's number if the customer withheld theirs. When true, they always see the assistant's own phone number, which is the number the customer called, so they can tell the call is a transfer. The tradeoff is that the customer's number is then not shown on their phone.

Tenant dictionaries to link, in precedence order. Replaces the whole set; [] unlinks all. A dictionary's own type decides whether its entries bias the transcriber or shape pronunciation.

Per-assistant STT keyterms and TTS pronunciation rules. Rules are provider-neutral: `phonemes` is space-separated IPA rendered per TTS model, `respell` is the plain-text fallback; at least one is required. Example: {"sttKeyterms":[{"term":"Gree","language":null}],"ttsPronunciations":[{"word":"Quatre","respell":"Kate","phonemes":"k a t","language":"fr"}]}. Replace semantics: send the whole object; null clears. Zod enforces the real shape.

IVR router configuration for this assistant. Free-form JSON object; no inner structure is validated yet while the feature is in progress. On PATCH this DEEP-MERGES into the stored document, so sending one key leaves the others as stored: no read-modify-write needed. Per-key removal is not expressible: a key sent as null is stored as null, which every consumer reads as unset. Send the whole field as null to clear the document (null = no IVR configured). Omit the field to leave it unchanged.