Call
Get phone number history across the organization
ModelsExpand Collapse
class CallRetrieveResponse: …
call: CallThis represent a call “order” that was requested by the user. A call order can be resolved over multiple call attempts spanning up to a few days.
This represent a call “order” that was requested by the user. A call order can be resolved over multiple call attempts spanning up to a few days.
assistant: Optional[CallAssistant]
Ambient background sound to play during the call. null disables it.
Volume of the ambient background sound (0 = silent, 1 = max).
calendly: Optional[CallAssistantCalendly]
The event type ID representing the event type to schedule. (eg: https://api.calendly.com/event_types/b2330295-2a91-4a1d-bb73-99e7707663d5)
call_retry_config: Optional[CallAssistantCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallAssistantCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
Optional Cartesia pronunciation dictionary ID linked to this assistant.
custom_tools: Optional[List[CallAssistantCustomTool]]
JSON body template for the request. Use quoted {{variable}} placeholders (e.g. “{{name}}”) for dynamic values
Human-readable description of what the tool does, used by the LLM to decide when to call it
headers: List[CallAssistantCustomToolHeader]HTTP headers to include in the request. Values support {{variable}} placeholders
HTTP headers to include in the request. Values support {{variable}} placeholders
input_schema: List[CallAssistantCustomToolInputSchema]Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Delay in milliseconds before speaking the first sentence. Default: 400.
Override the default outbound phone number for calls placed with this assistant. When null, the organization’s default phone number is used.
hubspot: Optional[CallAssistantHubspot]
meeting_links: List[CallAssistantHubspotMeetingLink]Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Public URL of the scheduling page (e.g. https://meetings-eu1.hubspot.com/julien). The agent books through the API and only texts this link to the caller when the booking fails and the call can receive an SMS. Null on calendars saved before this field existed.
Meeting duration in milliseconds, chosen among the durations the scheduling page offers (e.g. 1800000 for 30 minutes).
Marks the round-robin calendar the agent falls back to when the requested rep has no slot. At most one meeting link may set this.
Resolve the caller against HubSpot contacts before answering an inbound call, filling the contact_* prompt variables (contact_name, contact_owner, contact_lead_status, contact_has_open_deal). Off by default: it costs API calls before the phone is picked up.
post_call_sync: Optional[CallAssistantHubspotPostCallSync]Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Records calls in HubSpot after they end: the contact is created or updated from the phone number and the call is logged on its timeline. By default every call a person answered is recorded. When the post-call prompt talks about the CRM (in plain words or by mentioning the crm_push_call tool), the post-call agent decides instead which calls are recorded and what the logged call says.
human_transfer_mode: Optional[Literal["warm", "cold"]]Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Enable IVR navigation tools. When enabled, the assistant can send DTMF tones and skip turns to navigate phone menus.
When true, the agent’s noise-cancellation filter is Krisp VIVA Tel instead of the legacy BVCTelephony.
llm_model: CallAssistantLlmModel
class CallAssistantLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Public URL of the brand logo shown on the assistant’s demo page. Null when unset.
The maximum duration of the call in seconds. This is the maximum time the call will be allowed to run.
Optional message the agent will say, without being interruptible, when the call reaches its max duration. Kept short so it fits inside the farewell buffer. If null, the call ends silently.
precall_lookup_headers: Optional[List[CallAssistantPrecallLookupHeader]]HTTP headers sent with the pre-call lookup request. Null when none are configured.
HTTP headers sent with the pre-call lookup request. Null when none are configured.
URL fetched before an inbound call is answered to resolve the assistant’s prompt variables. Null when no lookup is configured.
slack: Optional[CallAssistantSlack]
structured_output_config: Optional[List[CallAssistantStructuredOutputConfig]]The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
Custom prompt for structured data extraction. If not provided, a default prompt is used. Available variables: {{transcript}}, {{call_direction}}, {{user_phone_number}}, {{agent_phone_number}}.
When false, the agent transcribes the raw caller track instead of the BVC noise-cancellation output.
stt_context: Optional[CallAssistantSttContext]Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Soniox endpoint latency adjustment level (0 to 3). Null: Soniox default.
Soniox endpoint sensitivity (-1 to 1). Null: Soniox default.
stt_model: Literal["stt-rt-v4", "stt-rt-v5"]Transcriber (speech-to-text) model used for the assistant.
Transcriber (speech-to-text) model used for the assistant.
teams: Optional[CallAssistantTeams]
channels: List[CallAssistantTeamsChannel]The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
thinking_sound: Optional[Literal["city-ambience.ogg", "forest-ambience.ogg", "office-ambience.ogg", 4 more]]Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Probability [0..1] that the thinking sound plays on any given turn; otherwise the agent is silent while thinking.
transfer_destinations: Optional[List[CallAssistantTransferDestination]]Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
voice: Optional[CallAssistantVoice]
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
voice_languages: Optional[List[CallAssistantVoiceLanguage]]Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
voice: CallAssistantVoiceLanguageVoiceThe voice used for contacts whose language is this language.
The voice used for contacts whose language is this language.
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
If set, when voicemail is detected the agent will speak this message then hang up; if null, hang up immediately.
Warm transfer only: instructions for the supervisor handoff summary; null when not configured or cold transfer.
Warm transfer only: voice speed of the briefing to the human colleague; null when it follows voice.speed, not configured, or cold transfer.
zoho: Optional[CallAssistantZoho]
field_mapping: List[CallAssistantZohoFieldMapping]Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
When true, also log the call as a record in Zoho’s Calls module (related to the upserted prospect via Who_Id).
Optional Note body template. Supports {{summary}}, {{outcome}}, {{phone}}, {{call_url}}, {{assistant_name}}, {{campaign_name}}, {{transcript}}, plus prompt_variables and structured_output keys. When null/empty a default note (outcome + summary + auto-listed extracted fields) is used.
call_attempts: List[CallCallAttempt]All call attempts for this call order, ordered by most recent first.
All call attempts for this call order, ordered by most recent first.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
call_retry_config: Optional[CallCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
channel: Literal["phone", "web"]web for a test call taken in the browser: its phone number is a placeholder, never dialed.
web for a test call taken in the browser: its phone number is a placeholder, never dialed.
contact: CallContactMagic contact variables (prospect identity) extracted from the call’s input data and transcript.
Magic contact variables (prospect identity) extracted from the call’s input data and transcript.
Delay in milliseconds before speaking the first sentence. Default: 400.
The phone number that made the call. Formatted in E.164 format. Example: +1234567890
DEPRECATED: Whether the call has been cancelled. This is derived from status. Use status instead.
DEPRECATED: Whether the call has completed. This is derived from status. Use status instead.
last_call_attempt: Optional[CallLastCallAttempt]This represent a single call attempt. A call attempt is a single call made to the phone number.
This represent a single call attempt. A call attempt is a single call made to the phone number.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallLastCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallLastCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallLastCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallLastCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallLastCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallLastCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallLastCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
llm_model: CallLlmModel
class CallLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Whether no further attempt will be made for this call: it was cancelled, or its future attempts were stopped while an attempt was already in progress (that attempt still finishes). Unlike is_cancelled, this is true as soon as retries are stopped, before the call reaches a terminal status.
The phone number that received the call. Formatted in E.164 format. Example: +1234567890
transfer_destinations: Optional[List[CallTransferDestination]]Where this call can be transferred, if anywhere.
Where this call can be transferred, if anywhere.
The time the call order was last updated (any state change, retry, or analysis result).
callback_from: Optional[CallCallbackFrom]For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
Why the call was cancelled, when it can be told. prospect_called_back: the prospect called back before the next retry, so the pending retry was cancelled and their call is recorded as a separate inbound call. Null for a manual or campaign cancel, and for calls that are not cancelled.
True when the prospect’s phone number is on the do-not-call list: it can no longer be called.
outcome: Optional[Literal["not_interested", "interested", "completed", "requested_callback_later"]]Business outcome of the call. Null if not computed or no transcript.
Business outcome of the call. Null if not computed or no transcript.
LLM explanation for the outcome, when outcome was computed from the transcript.
prompt_version: Optional[CallPromptVersion]The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.
The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.
class CallCreateResponse: …
call: CallThis represent a call “order” that was requested by the user. A call order can be resolved over multiple call attempts spanning up to a few days.
This represent a call “order” that was requested by the user. A call order can be resolved over multiple call attempts spanning up to a few days.
assistant: Optional[CallAssistant]
Ambient background sound to play during the call. null disables it.
Volume of the ambient background sound (0 = silent, 1 = max).
calendly: Optional[CallAssistantCalendly]
The event type ID representing the event type to schedule. (eg: https://api.calendly.com/event_types/b2330295-2a91-4a1d-bb73-99e7707663d5)
call_retry_config: Optional[CallAssistantCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallAssistantCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
Optional Cartesia pronunciation dictionary ID linked to this assistant.
custom_tools: Optional[List[CallAssistantCustomTool]]
JSON body template for the request. Use quoted {{variable}} placeholders (e.g. “{{name}}”) for dynamic values
Human-readable description of what the tool does, used by the LLM to decide when to call it
headers: List[CallAssistantCustomToolHeader]HTTP headers to include in the request. Values support {{variable}} placeholders
HTTP headers to include in the request. Values support {{variable}} placeholders
input_schema: List[CallAssistantCustomToolInputSchema]Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Delay in milliseconds before speaking the first sentence. Default: 400.
Override the default outbound phone number for calls placed with this assistant. When null, the organization’s default phone number is used.
hubspot: Optional[CallAssistantHubspot]
meeting_links: List[CallAssistantHubspotMeetingLink]Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Public URL of the scheduling page (e.g. https://meetings-eu1.hubspot.com/julien). The agent books through the API and only texts this link to the caller when the booking fails and the call can receive an SMS. Null on calendars saved before this field existed.
Meeting duration in milliseconds, chosen among the durations the scheduling page offers (e.g. 1800000 for 30 minutes).
Marks the round-robin calendar the agent falls back to when the requested rep has no slot. At most one meeting link may set this.
Resolve the caller against HubSpot contacts before answering an inbound call, filling the contact_* prompt variables (contact_name, contact_owner, contact_lead_status, contact_has_open_deal). Off by default: it costs API calls before the phone is picked up.
post_call_sync: Optional[CallAssistantHubspotPostCallSync]Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Records calls in HubSpot after they end: the contact is created or updated from the phone number and the call is logged on its timeline. By default every call a person answered is recorded. When the post-call prompt talks about the CRM (in plain words or by mentioning the crm_push_call tool), the post-call agent decides instead which calls are recorded and what the logged call says.
human_transfer_mode: Optional[Literal["warm", "cold"]]Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Enable IVR navigation tools. When enabled, the assistant can send DTMF tones and skip turns to navigate phone menus.
When true, the agent’s noise-cancellation filter is Krisp VIVA Tel instead of the legacy BVCTelephony.
llm_model: CallAssistantLlmModel
class CallAssistantLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Public URL of the brand logo shown on the assistant’s demo page. Null when unset.
The maximum duration of the call in seconds. This is the maximum time the call will be allowed to run.
Optional message the agent will say, without being interruptible, when the call reaches its max duration. Kept short so it fits inside the farewell buffer. If null, the call ends silently.
precall_lookup_headers: Optional[List[CallAssistantPrecallLookupHeader]]HTTP headers sent with the pre-call lookup request. Null when none are configured.
HTTP headers sent with the pre-call lookup request. Null when none are configured.
URL fetched before an inbound call is answered to resolve the assistant’s prompt variables. Null when no lookup is configured.
slack: Optional[CallAssistantSlack]
structured_output_config: Optional[List[CallAssistantStructuredOutputConfig]]The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
Custom prompt for structured data extraction. If not provided, a default prompt is used. Available variables: {{transcript}}, {{call_direction}}, {{user_phone_number}}, {{agent_phone_number}}.
When false, the agent transcribes the raw caller track instead of the BVC noise-cancellation output.
stt_context: Optional[CallAssistantSttContext]Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Soniox endpoint latency adjustment level (0 to 3). Null: Soniox default.
Soniox endpoint sensitivity (-1 to 1). Null: Soniox default.
stt_model: Literal["stt-rt-v4", "stt-rt-v5"]Transcriber (speech-to-text) model used for the assistant.
Transcriber (speech-to-text) model used for the assistant.
teams: Optional[CallAssistantTeams]
channels: List[CallAssistantTeamsChannel]The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
thinking_sound: Optional[Literal["city-ambience.ogg", "forest-ambience.ogg", "office-ambience.ogg", 4 more]]Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Probability [0..1] that the thinking sound plays on any given turn; otherwise the agent is silent while thinking.
transfer_destinations: Optional[List[CallAssistantTransferDestination]]Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
voice: Optional[CallAssistantVoice]
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
voice_languages: Optional[List[CallAssistantVoiceLanguage]]Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
voice: CallAssistantVoiceLanguageVoiceThe voice used for contacts whose language is this language.
The voice used for contacts whose language is this language.
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
If set, when voicemail is detected the agent will speak this message then hang up; if null, hang up immediately.
Warm transfer only: instructions for the supervisor handoff summary; null when not configured or cold transfer.
Warm transfer only: voice speed of the briefing to the human colleague; null when it follows voice.speed, not configured, or cold transfer.
zoho: Optional[CallAssistantZoho]
field_mapping: List[CallAssistantZohoFieldMapping]Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
When true, also log the call as a record in Zoho’s Calls module (related to the upserted prospect via Who_Id).
Optional Note body template. Supports {{summary}}, {{outcome}}, {{phone}}, {{call_url}}, {{assistant_name}}, {{campaign_name}}, {{transcript}}, plus prompt_variables and structured_output keys. When null/empty a default note (outcome + summary + auto-listed extracted fields) is used.
call_attempts: List[CallCallAttempt]All call attempts for this call order, ordered by most recent first.
All call attempts for this call order, ordered by most recent first.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
call_retry_config: Optional[CallCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
channel: Literal["phone", "web"]web for a test call taken in the browser: its phone number is a placeholder, never dialed.
web for a test call taken in the browser: its phone number is a placeholder, never dialed.
contact: CallContactMagic contact variables (prospect identity) extracted from the call’s input data and transcript.
Magic contact variables (prospect identity) extracted from the call’s input data and transcript.
Delay in milliseconds before speaking the first sentence. Default: 400.
The phone number that made the call. Formatted in E.164 format. Example: +1234567890
DEPRECATED: Whether the call has been cancelled. This is derived from status. Use status instead.
DEPRECATED: Whether the call has completed. This is derived from status. Use status instead.
last_call_attempt: Optional[CallLastCallAttempt]This represent a single call attempt. A call attempt is a single call made to the phone number.
This represent a single call attempt. A call attempt is a single call made to the phone number.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallLastCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallLastCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallLastCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallLastCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallLastCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallLastCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallLastCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
llm_model: CallLlmModel
class CallLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Whether no further attempt will be made for this call: it was cancelled, or its future attempts were stopped while an attempt was already in progress (that attempt still finishes). Unlike is_cancelled, this is true as soon as retries are stopped, before the call reaches a terminal status.
The phone number that received the call. Formatted in E.164 format. Example: +1234567890
transfer_destinations: Optional[List[CallTransferDestination]]Where this call can be transferred, if anywhere.
Where this call can be transferred, if anywhere.
The time the call order was last updated (any state change, retry, or analysis result).
callback_from: Optional[CallCallbackFrom]For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
Why the call was cancelled, when it can be told. prospect_called_back: the prospect called back before the next retry, so the pending retry was cancelled and their call is recorded as a separate inbound call. Null for a manual or campaign cancel, and for calls that are not cancelled.
True when the prospect’s phone number is on the do-not-call list: it can no longer be called.
outcome: Optional[Literal["not_interested", "interested", "completed", "requested_callback_later"]]Business outcome of the call. Null if not computed or no transcript.
Business outcome of the call. Null if not computed or no transcript.
LLM explanation for the outcome, when outcome was computed from the transcript.
prompt_version: Optional[CallPromptVersion]The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.
The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.
class CallSearchResponse: …
calls: List[Call]
assistant: Optional[CallAssistant]
Ambient background sound to play during the call. null disables it.
Volume of the ambient background sound (0 = silent, 1 = max).
calendly: Optional[CallAssistantCalendly]
The event type ID representing the event type to schedule. (eg: https://api.calendly.com/event_types/b2330295-2a91-4a1d-bb73-99e7707663d5)
call_retry_config: Optional[CallAssistantCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallAssistantCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
Optional Cartesia pronunciation dictionary ID linked to this assistant.
custom_tools: Optional[List[CallAssistantCustomTool]]
JSON body template for the request. Use quoted {{variable}} placeholders (e.g. “{{name}}”) for dynamic values
Human-readable description of what the tool does, used by the LLM to decide when to call it
headers: List[CallAssistantCustomToolHeader]HTTP headers to include in the request. Values support {{variable}} placeholders
HTTP headers to include in the request. Values support {{variable}} placeholders
input_schema: List[CallAssistantCustomToolInputSchema]Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Schema defining the parameters the LLM should extract from the conversation to pass to this tool
Delay in milliseconds before speaking the first sentence. Default: 400.
Override the default outbound phone number for calls placed with this assistant. When null, the organization’s default phone number is used.
hubspot: Optional[CallAssistantHubspot]
meeting_links: List[CallAssistantHubspotMeetingLink]Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Scheduling pages the agent can book on during a call — typically one per sales rep plus a round-robin team page marked as the fallback.
Public URL of the scheduling page (e.g. https://meetings-eu1.hubspot.com/julien). The agent books through the API and only texts this link to the caller when the booking fails and the call can receive an SMS. Null on calendars saved before this field existed.
Meeting duration in milliseconds, chosen among the durations the scheduling page offers (e.g. 1800000 for 30 minutes).
Marks the round-robin calendar the agent falls back to when the requested rep has no slot. At most one meeting link may set this.
Resolve the caller against HubSpot contacts before answering an inbound call, filling the contact_* prompt variables (contact_name, contact_owner, contact_lead_status, contact_has_open_deal). Off by default: it costs API calls before the phone is picked up.
post_call_sync: Optional[CallAssistantHubspotPostCallSync]Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Sends call data back to HubSpot after the call. Off until set: connecting HubSpot to book meetings must not start writing to the CRM. Always present in responses; omit it on update to leave the current setting untouched.
Records calls in HubSpot after they end: the contact is created or updated from the phone number and the call is logged on its timeline. By default every call a person answered is recorded. When the post-call prompt talks about the CRM (in plain words or by mentioning the crm_push_call tool), the post-call agent decides instead which calls are recorded and what the logged call says.
human_transfer_mode: Optional[Literal["warm", "cold"]]Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Warm or cold transfer when transfer_destinations is non-empty; null when transfer is not configured.
Enable IVR navigation tools. When enabled, the assistant can send DTMF tones and skip turns to navigate phone menus.
When true, the agent’s noise-cancellation filter is Krisp VIVA Tel instead of the legacy BVCTelephony.
llm_model: CallAssistantLlmModel
class CallAssistantLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Public URL of the brand logo shown on the assistant’s demo page. Null when unset.
The maximum duration of the call in seconds. This is the maximum time the call will be allowed to run.
Optional message the agent will say, without being interruptible, when the call reaches its max duration. Kept short so it fits inside the farewell buffer. If null, the call ends silently.
precall_lookup_headers: Optional[List[CallAssistantPrecallLookupHeader]]HTTP headers sent with the pre-call lookup request. Null when none are configured.
HTTP headers sent with the pre-call lookup request. Null when none are configured.
URL fetched before an inbound call is answered to resolve the assistant’s prompt variables. Null when no lookup is configured.
slack: Optional[CallAssistantSlack]
structured_output_config: Optional[List[CallAssistantStructuredOutputConfig]]The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
The structured output config to use for the call. This is used to extract the data from the call (like email, name, company name, etc.).
Custom prompt for structured data extraction. If not provided, a default prompt is used. Available variables: {{transcript}}, {{call_direction}}, {{user_phone_number}}, {{agent_phone_number}}.
When false, the agent transcribes the raw caller track instead of the BVC noise-cancellation output.
stt_context: Optional[CallAssistantSttContext]Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Assistant-level speech-to-text context: structured general key/value pairs plus a list of domain terms. Prompt-derived context is merged in without replacing existing entries.
Soniox endpoint latency adjustment level (0 to 3). Null: Soniox default.
Soniox endpoint sensitivity (-1 to 1). Null: Soniox default.
stt_model: Literal["stt-rt-v4", "stt-rt-v5"]Transcriber (speech-to-text) model used for the assistant.
Transcriber (speech-to-text) model used for the assistant.
teams: Optional[CallAssistantTeams]
channels: List[CallAssistantTeamsChannel]The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
The Teams channels the post-call agent may post to. It resolves the channel it was told to use against this list and cannot reach any other.
thinking_sound: Optional[Literal["city-ambience.ogg", "forest-ambience.ogg", "office-ambience.ogg", 4 more]]Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Audio clip to play while the agent is processing a response. One of the built-in LiveKit audio clips; null disables it.
Probability [0..1] that the thinking sound plays on any given turn; otherwise the agent is silent while thinking.
transfer_destinations: Optional[List[CallAssistantTransferDestination]]Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
Where calls can be transferred to when users ask to speak to a human. Null or empty when transfer is not configured.
voice: Optional[CallAssistantVoice]
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
voice_languages: Optional[List[CallAssistantVoiceLanguage]]Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
Extra languages for a multilingual assistant, each with its own voice. Calls pick one through their language value, else the phone number’s country; other calls use the assistant’s voice.
voice: CallAssistantVoiceLanguageVoiceThe voice used for contacts whose language is this language.
The voice used for contacts whose language is this language.
model: Optional[Literal["sonic-3", "sonic-3.5", "sonic-3.6"]]Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
Cartesia TTS model (Cartesia only). Defaults to sonic-3.6 when omitted. Ignored for other providers.
If set, when voicemail is detected the agent will speak this message then hang up; if null, hang up immediately.
Warm transfer only: instructions for the supervisor handoff summary; null when not configured or cold transfer.
Warm transfer only: voice speed of the briefing to the human colleague; null when it follows voice.speed, not configured, or cold transfer.
zoho: Optional[CallAssistantZoho]
field_mapping: List[CallAssistantZohoFieldMapping]Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
Maps extracted call fields (structured_output / prompt variables) onto Zoho field API names on the upserted record.
When true, also log the call as a record in Zoho’s Calls module (related to the upserted prospect via Who_Id).
Optional Note body template. Supports {{summary}}, {{outcome}}, {{phone}}, {{call_url}}, {{assistant_name}}, {{campaign_name}}, {{transcript}}, plus prompt_variables and structured_output keys. When null/empty a default note (outcome + summary + auto-listed extracted fields) is used.
call_attempts: List[CallCallAttempt]All call attempts for this call order, ordered by most recent first.
All call attempts for this call order, ordered by most recent first.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
call_retry_config: Optional[CallCallRetryConfig]Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
Configuration for call retry behavior including time windows, delays, and max iterations. If not provided, defaults will be used.
allowed_days: List[Literal["monday", "tuesday", "wednesday", 4 more]]Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
Days of the week when calls are allowed, in the recipient’s timezone. Default: Monday through Friday.
If true and max_retry_attempts >= 2, attempt #2 fires immediately (skipping retry_delay_seconds) when attempt #1 didn’t reach a human. Calling-window/allowed-days checks still apply. Only affects the 1→2 transition. Default: false.
calling_windows: List[CallCallRetryConfigCallingWindow]
End time for the calling window in the recipient’s timezone (or timezone_override if provided). Format: ‘HH:mm’ (24-hour) or ‘H:mma’ (12-hour). Examples: ‘17:00’, ‘6pm’. Default: ‘18:00’.
channel: Literal["phone", "web"]web for a test call taken in the browser: its phone number is a placeholder, never dialed.
web for a test call taken in the browser: its phone number is a placeholder, never dialed.
contact: CallContactMagic contact variables (prospect identity) extracted from the call’s input data and transcript.
Magic contact variables (prospect identity) extracted from the call’s input data and transcript.
Delay in milliseconds before speaking the first sentence. Default: 400.
The phone number that made the call. Formatted in E.164 format. Example: +1234567890
DEPRECATED: Whether the call has been cancelled. This is derived from status. Use status instead.
DEPRECATED: Whether the call has completed. This is derived from status. Use status instead.
last_call_attempt: Optional[CallLastCallAttempt]This represent a single call attempt. A call attempt is a single call made to the phone number.
This represent a single call attempt. A call attempt is a single call made to the phone number.
dial_error: Optional[Literal["number_non_attributed", "too_many_calls", "busy", 7 more]]The SIP error that occurred.
The SIP error that occurred.
Whether the AI agent conducted the conversation on this attempt. False for a silent speed-dial probe (whose agent only detects a human) and for attempts that were never answered.
Whether a human conducted (part of) this attempt — a jump-in / speed-dial takeover, or a transfer to a rep.
handled_by_user: Optional[CallLastCallAttemptHandledByUser]The team member responsible for following up on this call.
The team member responsible for following up on this call.
actions: Optional[List[CallLastCallAttemptAction]]Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Tools the agent ran for this attempt, in call and after it, ordered oldest first. This is the decision trail — the delivery trail lives in sms_log / email_log.
Names of the assistants in assistants_used, keyed by id. Only set when the attempt went through several assistants (a multi-step call).
Assistant node ids entered during this attempt, in traversal order.
email_log: Optional[List[CallLastCallAttemptEmailLog]]Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
Notification emails sent in connection with the call, ordered oldest first (from/to addresses, subject, sent time).
end_reason: Optional[Literal["client_initiated", "connection_timeout", "human_takeover", 13 more]]Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
Reason for ending the call when ended_by is ‘agent’. E.g. ‘tool_end_call’, ‘voicemail’, ‘transfer’, ‘ivr_no_navigate’.
ended_by: Optional[Literal["agent", "user", "system"]]Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
Who ended the call: ‘agent’ (AI agent), ‘user’ (caller/callee hung up), or ‘system’ (e.g. max duration limit).
post_call_transcript: Optional[List[CallLastCallAttemptPostCallTranscript]]Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
Higher-quality transcript generated after the call via Soniox async STT. Null until generated or when no recording is available.
sms_log: Optional[List[CallLastCallAttemptSMSLog]]Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
Automatic SMS sent after the call (voicemail and after-call), ordered oldest first. The in-call send_sms is excluded — it already appears in the transcript as a tool call.
The provider’s numeric reason for a terminal failure (e.g. Twilio 30034). Null unless the message failed.
speech_events: Optional[List[CallLastCallAttemptSpeechEvent]]Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Per-turn speech segments (VAD-derived) with timing and pipeline latencies, ordered by started_speaking_at_ms. Empty for calls recorded before this was tracked or for realtime sessions that don’t emit them.
Agent-only: despite the name, the whole prompt the LLM was given for this turn — system instructions plus conversation history — as JSON.
Agent-only: the model that produced the reply — the winning leg when the LLM is a race, which is also the model llm_ttft_ms describes.
Milliseconds since session start when speech ended. Null when the session ended before the turn closed.
The data extracted from the call, using the structured output config from the parent call object.
transcript: Optional[List[CallLastCallAttemptTranscript]]The live transcript of the call, built in real time.
The live transcript of the call, built in real time.
The number the call was transferred to, in E.164 format (e.g. +33123456789). Recorded by the agent from the destination it actually dialled, so it is the destination itself and not a name read off the transcript. Null when the call was not transferred, and on transfers placed before this was tracked.
llm_model: CallLlmModel
class CallLlmModelUnionMember5: …
Fireworks model id, e.g. accounts/fireworks/models/qwen3-235b-a22b or a dedicated deployment id.
Fireworks AI model, served with the server-side FIREWORKS_API_KEY (no key stored per-assistant).
reasoning_effort: Optional[Literal["none", "low", "medium", "high"]]Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Reasoning effort sent with each request. Hybrid-reasoning models (e.g. gemma-4-31b) think by default, which adds latency and can eat the whole token budget before any speakable text: voice agents want ‘none’. Omit for models that reject the knob.
Whether no further attempt will be made for this call: it was cancelled, or its future attempts were stopped while an attempt was already in progress (that attempt still finishes). Unlike is_cancelled, this is true as soon as retries are stopped, before the call reaches a terminal status.
The phone number that received the call. Formatted in E.164 format. Example: +1234567890
transfer_destinations: Optional[List[CallTransferDestination]]Where this call can be transferred, if anywhere.
Where this call can be transferred, if anywhere.
The time the call order was last updated (any state change, retry, or analysis result).
callback_from: Optional[CallCallbackFrom]For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
For an inbound call only: the outbound call this one answers, when the prospect rang us back after we called them. Null for every other call.
Why the call was cancelled, when it can be told. prospect_called_back: the prospect called back before the next retry, so the pending retry was cancelled and their call is recorded as a separate inbound call. Null for a manual or campaign cancel, and for calls that are not cancelled.
True when the prospect’s phone number is on the do-not-call list: it can no longer be called.
outcome: Optional[Literal["not_interested", "interested", "completed", "requested_callback_later"]]Business outcome of the call. Null if not computed or no transcript.
Business outcome of the call. Null if not computed or no transcript.
LLM explanation for the outcome, when outcome was computed from the transcript.
prompt_version: Optional[CallPromptVersion]The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.
The assistant prompt version the call’s prompt was compiled from. Null for calls created before versions were tracked, and for calls made with an inline assistant.