---
title: "POST /v3/create-web-call"
method: POST
path: "/v3/create-web-call"
---

# POST /v3/create-web-call

`POST /v3/create-web-call`

## Request

- Base URL: `https://api.retellai.com`
- URL: `https://api.retellai.com/v3/create-web-call`
- Auth: HTTP bearer

Create a new web call and return browser connection details.

## Request body

- object
  - `agent_id` string, required: Unique id of agent used for the call. Your agent would contain the LLM Websocket url used for this call.
  - `agent_version` union: Agent version reference. Supports a numeric version (for example 3) or a tag/environment name (for example "prod"). The string "latest" resolves to the most recently created version (the largest version number), and "latest_published" resolves to the most recently published version. When a tag is provided, resolution uses that exact tag assignment (including its dynamic variables). If the tag exists but is currently unassigned, it resolves to latest. When a numeric version, latest, or latest_published is provided, resolution applies dynamic variables from the preferred tag for that resolved version (most recently assigned), if any.
    - string
    - integer
  - `agent_override` AgentOverrideRequest: Override configuration for agent, retell LLM, or conversation flow settings for a specific call.
    - `agent` AgentRequest
      - `response_engine` union
        - ResponseEngineRetellLm
          - `type` 'retell-llm', required: type of the Response Engine.
          - `llm_id` string, required: id of the Retell LLM Response Engine.
          - `version` number, nullable: Version of the Retell LLM Response Engine.
        - ResponseEngineCustomLm
          - `type` 'custom-llm', required: type of the Response Engine.
          - `llm_websocket_url` string, required: LLM websocket url of the custom LLM.
        - ResponseEngineConversationFlow
          - `type` 'conversation-flow', required: type of the Response Engine.
          - `conversation_flow_id` string, required: ID of the Conversation Flow Response Engine.
          - `version` number, nullable: Version of the Conversation Flow Response Engine.
      - `agent_name` string, nullable: The name of the agent. Only used for your own reference.
      - `version_description` string, nullable: Optional description of the agent version. Used for your own reference and documentation.
      - `version_title` string, nullable: Optional title of the agent version. Used for your own reference.
      - `voice_id` string: Unique voice id used for the agent. Find list of available voices and their preview in Dashboard.
      - `voice_model` 'eleven_flash_v2' | 'eleven_flash_v2_5' | 'eleven_multilingual_v2' | 'eleven_v3' | 'eleven_v3_conversational' | 'eleven_v4' | 'eleven_v4_turbo' | 'sonic-3' | 'sonic-3-latest' | 'sonic-3.5' | 'sonic-3.6' | 'tts-1' | 'gpt-4o-mini-tts' | 'speech-02-turbo' | 'speech-2.8-turbo' | 's1' | 's2-pro' | 's2.1-pro' | 'drama-3' | 'inworld-tts-2' | 'inworld-tts-2-flash' | 'null', nullable: Select the voice model used for the selected voice. Each provider has a set of available voice models. Set to null to remove voice model selection, and default ones will apply. Check out dashboard for more details of each voice model.
      - `fallback_voice_ids` string[], nullable: When TTS provider for the selected voice is experiencing outages, we would use fallback voices listed here for the agent. Voice id and the fallback voice ids must be from different TTS providers. The system would go through the list in order, if the first one in the list is also having outage, it would use the next one. Set to null to remove voice fallback for the agent.
      - `voice_temperature` number: Controls how stable the voice is. Value ranging from [0,2]. Lower value means more stable, and higher value means more variant speech generation. Check the dashboard to see what provider supports this feature. If unset, default value 1 will apply.
      - `voice_speed` number: Controls speed of voice. Value ranging from [0.5,2]. Lower value means slower speech, while higher value means faster speech rate. If unset, default value 1 will apply.
      - `enable_dynamic_voice_speed` boolean: If set to true, will enable dynamic voice speed adjustment based on the user's speech rate and conversation context. If unset, default value false will apply.
      - `enable_dynamic_responsiveness` boolean: If set to true, the agent will dynamically adjust how quickly it responds based on the user's speech rate and past turn-taking behavior in the call. If unset, default value false will apply.
      - `volume` number: If set, will control the volume of the agent. Value ranging from [0,2]. Lower value means quieter agent speech, while higher value means louder agent speech. If unset, default value 1 will apply.
      - `enable_expressive_mode` boolean: Master toggle for expressive mode. When true, the agent may add expressive voice tags to the audio it generates. Only applicable for platform voices. If unset, defaults to false.
      - `expressive_emotion_tags` string[]: The expressive voice tags Retell pre-teaches the model to use when enable_expressive_mode is true. Custom tags defined in the system prompt are still allowed. If empty, the agent follows general expressive guidance without a fixed tag set.
      - `expressive_mode_prompt` string, nullable: Custom expressive voice guidance to use instead of the default Retell expressive prompt when enable_expressive_mode is true. If omitted or blank, the default expressive prompt will be used.
      - `responsiveness` number: Controls how responsive is the agent. Value ranging from [0,1]. Lower value means less responsive agent (wait more, respond slower), while higher value means faster exchanges (respond when it can). If unset, default value 1 will apply.
      - `interruption_sensitivity` number: Controls how sensitive the agent is to user interruptions. Value ranging from [0,1]. Lower value means it will take longer / more words for user to interrupt agent, while higher value means it's easier for user to interrupt agent. If unset, default value 1 will apply. When this is set to 0, agent would never be interrupted.
      - `enable_backchannel` boolean: Controls whether the agent would backchannel (agent interjects the speaker with phrases like "yeah", "uh-huh" to signify interest and engagement). Backchannel when enabled tends to show up more in longer user utterances. If not set, agent will not backchannel.
      - `backchannel_frequency` number: Only applicable when enable_backchannel is true. Controls how often the agent would backchannel when a backchannel is possible. Value ranging from [0,1]. Lower value means less frequent backchannel, while higher value means more frequent backchannel. If unset, default value 0.8 will apply.
      - `backchannel_words` string[], nullable: Only applicable when enable_backchannel is true. A list of words that the agent would use as backchannel. If not set, default backchannel words will apply. Check out [backchannel default words](/agent/interaction-configuration#backchannel) for more details. Note that certain voices do not work too well with certain words, so it's recommended to experiment before adding any words.
      - `reminder_trigger_ms` number: If set (in milliseconds), will trigger a reminder to the agent to speak if the user has been silent for the specified duration after some agent speech. Must be a positive number. If unset, default value of 10000 ms (10 s) will apply.
      - `reminder_max_count` integer: If set, controls how many times agent would remind user when user is unresponsive. Must be a non negative integer. If unset, default value of 1 will apply (remind once). Set to 0 to disable agent from reminding.
      - `ambient_sound` 'coffee-shop' | 'convention-hall' | 'summer-outdoor' | 'mountain-outdoor' | 'static-noise' | 'call-center' | 'null', nullable: If set, will add ambient environment sound to the call to make experience more realistic. Currently supports the following options: - `coffee-shop`: Coffee shop ambience with people chatting in background. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/coffee-shop.wav) - `convention-hall`: Convention hall ambience, with some echo and people chatting in background. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/convention-hall.wav) - `summer-outdoor`: Summer outdoor ambience with cicada chirping. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/summer-outdoor.wav) - `mountain-outdoor`: Mountain outdoor ambience with birds singing. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/mountain-outdoor.wav) - `static-noise`: Constant static noise. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/static-noise.wav) - `call-center`: Call center work noise. [Listen to Ambience](https://retell-utils-public.s3.us-west-2.amazonaws.com/call-center.wav) Set to `null` to remove ambient sound from this agent.
      - `ambient_sound_volume` number: If set, will control the volume of the ambient sound. Value ranging from [0,2]. Lower value means quieter ambient sound, while higher value means louder ambient sound. If unset, default value 1 will apply.
      - `language` union: Specifies what language(s) the agent will operate in. Accepts either a single locale (e.g. `en-US`) or an array of locales for multilingual agents (e.g. `["en-US","es-ES"]`). If unset, defaults to `en-US`.
        - 'en-US' | 'en-IN' | 'en-GB' | 'en-AU' | 'en-NZ' | 'de-DE' | 'es-ES' | 'es-419' | 'hi-IN' | 'fr-FR' | 'fr-CA' | 'ja-JP' | 'pt-PT' | 'pt-BR' | 'zh-CN' | 'ru-RU' | 'it-IT' | 'ko-KR' | 'nl-NL' | 'nl-BE' | 'pl-PL' | 'tr-TR' | 'vi-VN' | 'ro-RO' | 'bg-BG' | 'ca-ES' | 'th-TH' | 'da-DK' | 'fi-FI' | 'el-GR' | 'hu-HU' | 'id-ID' | 'no-NO' | 'sk-SK' | 'sv-SE' | 'lt-LT' | 'lv-LV' | 'cs-CZ' | 'ms-MY' | 'af-ZA' | 'ar-SA' | 'az-AZ' | 'bs-BA' | 'cy-GB' | 'fa-IR' | 'fil-PH' | 'gl-ES' | 'he-IL' | 'hr-HR' | 'hy-AM' | 'is-IS' | 'kk-KZ' | 'kn-IN' | 'mk-MK' | 'mr-IN' | 'ne-NP' | 'sl-SI' | 'sr-RS' | 'sw-KE' | 'ta-IN' | 'ur-IN' | 'yue-CN' | 'uk-UA': Specifies what language (and dialect) the agent will operate in. For instance, selecting `en-GB` optimizes speech recognition for British English and indexes knowledge bases with English. If unset, will use default value `en-US`.
        - Language[]
      - `webhook_url` string, nullable: The webhook for agent to listen to call events. See what events it would get at [webhook doc](/features/webhook). If set, will binds webhook events for this agent to the specified url, and will ignore the account level webhook for this agent. Set to `null` to remove webhook url from this agent.
      - `webhook_events` string[], nullable: Which webhook events this agent should receive. If not set, defaults to call_started, call_ended, call_analyzed.
      - `webhook_timeout_ms` integer: The timeout for the webhook in milliseconds. If not set, default value of 10000 will apply.
      - `boosted_keywords` string[], nullable: Provide a customized list of keywords to bias the transcriber model, so that these words are more likely to get transcribed. Commonly used for names, brands, street, etc. Entries may reference dynamic variables with `{{variable}}` syntax.
      - `contact_memory_config` ContactMemoryConfig: Contact memory settings for phone calls and SMS chats. Creating an agent defaults enable_update to false and enable_read to true. Updates only change the supplied flags; omitted flags stay unchanged and an empty object has no effect. Set a flag to false to disable it. The configuration cannot be cleared. Existing agents without this configuration have both disabled.
        - `enable_update` boolean: Rewrite the contact memory after each conversation. Cannot be enabled for payment agents. Requires storing conversation data. Chat agents must also have end_chat_after_silence_ms set.
        - `enable_read` boolean: Automatically add saved contact memory to the agent prompt. Skippable nodes can use answers from the current conversation even when this setting is disabled. Contact dynamic variables, including contact_memory, remain available regardless of this setting.
      - `data_storage_setting` 'everything' | 'everything_except_pii' | 'basic_attributes_only': Granular setting to manage how Retell stores sensitive data (transcripts, recordings, logs, etc.). This replaces the deprecated `opt_out_sensitive_data_storage` field. - `everything`: Store all data including transcripts, recordings, and logs. - `everything_except_pii`: Store data without PII when PII is detected. - `basic_attributes_only`: Store only basic attributes; no transcripts/recordings/logs. If not set, default value of "everything" will apply.
      - `data_storage_retention_days` integer, nullable: Number of days to retain call/chat data before automatic deletion. Must be between 1 and 730 days. If not set, data is retained forever (no automatic deletion).
      - `opt_in_signed_url` boolean: Whether this agent opts in for signed URLs for public logs and recordings. When enabled, the generated URLs will include security signatures that restrict access and automatically expire after 24 hours.
      - `signed_url_expiration_ms` integer, nullable: The expiration time for the signed url in milliseconds. Only applicable when opt_in_signed_url is true. If not set, default value of 86400000 (24 hours) will apply.
      - `pronunciation_dictionary` object[], nullable: A list of words / phrases and their pronunciation to be used to guide the audio synthesize for consistent pronunciation. Check the dashboard to see what provider supports this feature. Set to null to remove pronunciation dictionary from this agent.
        - `word` string, required: The string of word / phrase to be annotated with pronunciation.
        - `alphabet` 'ipa' | 'cmu' | 'pinyin' | 'jyutping', required: The phonetic alphabet to use. MiniMax speech-02-turbo supports IPA and Pinyin. MiniMax speech-2.8-turbo also supports Jyutping. Support for other alphabets depends on the selected voice provider and model.
        - `phoneme` string, required: Pronunciation of the word in the format of the selected phonetic alphabet.
      - `enable_dnc_detection` boolean: If set to true, the agent recognizes requests to stop calling or contacting the user, confirms once, and on a clear yes ends the call with disconnection reason user_requested_dnc and sets do_not_call to true on the contact for the user's phone number. If unset, default value false will apply.
      - `end_call_after_silence_ms` integer: If users stay silent for a period after agent speech, end the call. The minimum value allowed is 10,000 ms (10 s). By default, this is set to 600000 (10 min).
      - `max_call_duration_ms` integer: Maximum allowed length for the call, will force end the call if reached. The minimum value allowed is 60,000 ms (1 min), and maximum value allowed is 7,200,000 (2 hours). By default, this is set to 3,600,000 (1 hour).
      - `voicemail_option` object, nullable: If this option is set, the call will try to detect voicemail in the first 3 minutes of the call. Actions defined (hangup, or leave a message) will be applied when the voicemail is detected. Set this to null to disable voicemail detection.
        - `action` union, required
          - VoicemailActionPrompt
            - `type` 'prompt', required
            - `text` string, required: The prompt used to generate the text to be spoken when the call is detected to be in voicemail.
          - VoicemailActionStaticText
            - `type` 'static_text', required
            - `text` string, required: The text to be spoken when the call is detected to be in voicemail.
          - VoicemailActionHangup
            - `type` 'hangup', required
          - VoicemailActionBridgeTransfer
            - `type` 'bridge_transfer', required
        - `detection_prompt` string, nullable: Optionally describe what should be treated as voicemail. Leave as null to use the default definition.
      - `ivr_option` object, nullable: If this option is set, the call will try to detect IVR in the first 3 minutes of the call. Actions defined will be applied when the IVR is detected. Set this to null to disable IVR detection.
        - `action` IvrAction, required
          - `type` 'hangup', required
        - `detection_prompt` string, nullable: Optionally describe what should be treated as an IVR. Leave as null to use the default definition.
      - `call_screening_option` CallScreeningOption, nullable: If this option is set, the agent prompt will include call screen handling instructions for identity and call purpose questions. Set this to null to disable call screen prompt instructions.
        - `agent_identity` string, required: Identity the agent should provide when a call screen asks who is calling. Dynamic variables are supported.
        - `call_purpose` string, required: Purpose the agent should provide when a call screen asks why it is calling. Dynamic variables are supported.
      - `post_call_analysis_data` PostCallAnalysisData[], nullable: Post call analysis data to extract from the call. This data will augment the pre-defined variables extracted in the call analysis. This will be available after the call ends.
        - union: Post-call analysis item (custom data or voice preset). Use for voice agent post_call_analysis_data; validates only call presets (call_summary, call_successful, user_sentiment).
          - union
            - StringAnalysisData
              - …
            - EnumAnalysisData
              - …
            - BooleanAnalysisData
              - …
            - NumberAnalysisData
              - …
          - CallPresetAnalysisData: System preset for post-call analysis (voice agents). Use in post_call_analysis_data to override prompts or mark fields optional.
            - `type` 'system-presets', required: Identifies this item as a system preset.
            - `name` 'call_summary' | 'call_successful' | 'user_sentiment', required: Preset identifier for voice agent analysis.
            - `description` string: Prompt or description for this preset.
            - `required` boolean: If false, this field is optional in the analysis. If true or unset, the field is required.
            - `conditional_prompt` string: Optional instruction to help decide whether this field needs to be populated. If not set, the field is always included.
      - `post_call_analysis_model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon' | 'null', nullable: Available LLM models for agents.
      - `begin_message_delay_ms` integer: If set, will delay the first message by the specified amount of milliseconds, so that it gives user more time to prepare to take the call. Valid range is [0, 5000]. If not set or set to 0, agent will speak immediately. Only applicable when agent speaks first.
      - `ring_duration_ms` integer: If set, the phone ringing will last for the specified amount of milliseconds. This applies for both outbound call ringtime, and call transfer ringtime. Default to 30000 (30 s). Valid range is [5000, 300000].
      - `stt_mode` 'fast' | 'accurate' | 'custom': If set, determines whether speech to text should focus on latency or accuracy. Default to fast mode. When set to custom, custom_stt_config must be provided.
      - `custom_stt_config` object, nullable: Custom STT configuration. Only used when stt_mode is set to custom.
        - `provider` 'azure' | 'deepgram' | 'soniox' | 'assemblyai' | 'muse', required: ASR provider name.
        - `endpointing_ms` integer, required: Endpointing timeout in milliseconds. Minimum is 100 for Azure, 10 for Deepgram, 500 for Soniox, 100 for AssemblyAI, 100 for Muse. For AssemblyAI, this sets min_turn_silence (100-3000 ms). max_turn_silence adds half of this value, rounded to the nearest millisecond and bounded to 500-1000 ms, with a total cap of 3000 ms. Muse detects turn ends itself and ignores this value.
      - `vocab_specialization` 'general' | 'medical': If set, determines the vocabulary set to use for transcription. This setting only applies for English agents, for non English agent, this setting is a no-op. Default to general.
      - `allow_user_dtmf` boolean: If set to true, DTMF input will be accepted and processed. If false, any DTMF input will be ignored. Default to true.
      - `allow_dtmf_interruption` boolean: If set to true, DTMF input will interrupt the agent even when interruption_sensitivity is 0. Can be overridden per conversation or subagent node. Default to false.
      - `user_dtmf_options` object, nullable
        - `digit_limit` number, nullable: The maximum number of digits allowed in the user's DTMF (Dual-Tone Multi-Frequency) input per turn. Once this limit is reached, the input is considered complete and a response will be generated immediately.
        - `termination_key` string, nullable: A single key that signals the end of DTMF input. Acceptable values include any digit (0-9), the pound/hash symbol (#), or the asterisk (*).
        - `timeout_ms` integer: The time (in milliseconds) to wait for user DTMF input before timing out. The timer resets with each digit received.
      - `denoising_mode` 'no-denoise' | 'noise-cancellation' | 'noise-and-background-speech-cancellation': If set, determines what denoising mode to use. Use "no-denoise" to bypass all audio denoising. Default to noise-cancellation.
      - `denoising_enhancement_level` number, nullable: Controls the enhancement level for background voice cancellation. Set to 0 to bypass background voice cancellation without BVC charges. Value ranging from [0,1]. Only applicable when denoising_mode is noise-and-background-speech-cancellation. Defaults to 0.8 if no value is configured. Set to null to clear the configured value. Omitting this field preserves the existing value.
      - `pii_config` PIIConfig
        - `mode` 'post_call', required: The processing mode for PII scrubbing. Currently only post-call is supported.
        - `categories` string[], required: List of PII categories to scrub from transcripts and recordings. PII redaction is only active when this list is non-empty; an empty array means no PII scrubbing is performed.
      - `guardrail_config` GuardrailConfig
        - `output_topics` string[], nullable: Selected prohibited agent topic categories to check. When agent messages contain these topics, they will be replaced with a placeholder message.
        - `input_topics` string[], nullable: Selected prohibited user topic categories to check. When user messages contain these topics, the agent will respond with a placeholder message instead of processing the request.
      - `handbook_config` VoiceHandbookConfig: Behavior presets for voice agents. All presets are available.
        - `default_personality` boolean: Professional call center rep baseline.
        - `conversational_personality` boolean: Enables Conversational Personality. When true, the agent uses the Conversational Personality handbook preset, skips Professional Rep Personality during prompt assembly, and enables internal colloquial rewrite behavior.
        - `natural_filler_words` boolean: Sprinkle natural speech fillers like "um", "you know" for a more human, conversational tone.
        - `high_empathy` boolean: Warm acknowledgment of caller concerns.
        - `echo_verification` boolean: Repeat back and confirm important details (voice only).
        - `nato_phonetic_alphabet` boolean: Spell using NATO phonetic alphabet style (voice only).
        - `speech_normalization` boolean: Convert numbers/dates/currency to spoken forms (voice only).
        - `smart_matching` boolean: Treat near-match similar words as same entity to reduce impact of transcription error (voice only).
        - `ai_disclosure` boolean: When asked, acknowledge being a virtual assistant.
        - `scope_boundaries` boolean: Stay within prompt/context scope, don't invent details.
      - `timezone` string, nullable: IANA timezone for the agent (e.g. America/New_York). Defaults to America/Los_Angeles if not set.
      - `pre_session_tools` PreSessionTool[], nullable: Integration (Agent Functions) tools run as a dependency graph during session setup, before the agent's first message. Outputs are injected as dynamic variables. On calls the graph gets one minute unless an outbound caller will be dialed after setup, in which case it gets five minutes. Past that session initialization continues and any remaining tools finish in the background, so their outputs no longer reach the agent's prompt. Set to null to clear.
        - union
          - object: A pre-session tool (app / custom / code, discriminated by type) plus its dependency edges. The tool's name is the depends_on handle and must be unique across the agent's session tools. Parameters may set a constant value or a description for the LLM to infer the value. Pre-session inference uses the agent prompt and dynamic variables.
            - `type` 'integration_app', required
            - `name` string, required: Name of the tool. Must be unique within the phase's tools; referenced by depends_on. Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `app_id` string, required: The connection (App) this tool runs against. Must be a connection in the organization whose provider matches this tool's provider.
            - `provider` string, required: Provider of the connection. Must match the connection's provider; supported providers are listed by list-app-templates.
            - `app_tool_template_name` string, required: Name of the catalog template within the provider, as listed by list-app-templates.
            - `parameters` ToolParameter[]: The resolved input parameters, in order. Properties may pin a value with const (including {{variable}} references) or provide a description for LLM inference. Each property may also record selected_input_mode, the editor mode the user selected ("const_enum", "const_boolean", "const_value", "description_custom", or "description_preset"); it is stored and returned as-is, used only by the tool config UI. Omit the key when no mode is recorded; when set, const_* modes require a non-empty const, and description_* modes must omit const entirely. Each parameter's required list must match the schema returned by the corresponding step of the get-app-tool-schema loop.
              - …
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: Mapping of a dynamic-variable name to the response field (dot-path) it is populated from. Missing paths are ignored.
            - `output_selection` union: Which fields of a successful JSON tool response the agent and the transcript see. Omit to keep all fields. Does not affect response_variables, which are always extracted from the raw response. Non-JSON responses and failed tool results are not filtered. Response size limits still apply.
              - …
            - `description` string: Only applies to during conversation functions; ignored by the pre/post conversation. Overrides the catalog template's LLM-facing description.
            - `speak_during_execution` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. If true, will speak during execution.
            - `speak_after_execution` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. Determines whether the agent would call LLM another time and speak when the result of the tool is obtained.
            - `execution_message_description` string: Only applies to during conversation functions; ignored by the pre/post conversation. The message for the agent to speak when executing the tool. Only applicable when speak_during_execution is true.
            - `execution_message_type` 'prompt' | 'static_text': Only applies to during conversation functions; ignored by the pre/post conversation. Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `enable_typing_sound` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. If true, play a typing sound on the agent audio track while this tool is executing. Useful when the tool takes a noticeable amount of time to prevent silence on the call.
            - `depends_on` string[]: Names of tools that must run before this one.
          - object: A pre-session tool (app / custom / code, discriminated by type) plus its dependency edges. The tool's name is the depends_on handle and must be unique across the agent's session tools. Parameters may set a constant value or a description for the LLM to infer the value. Pre-session inference uses the agent prompt and dynamic variables.
            - `type` 'custom', required
            - `name` string, required: Name of the tool. Must be unique within all tools available to LLM at any given time (general tools + state tools + state edges). Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `url` string, required: Describes what the tool does, sometimes can also include information about when to call the tool.
            - `description` string: Describes what this tool does and when to call this tool.
            - `method` 'GET' | 'POST' | 'PUT' | 'PATCH' | 'DELETE': Method to use for the request, default to POST.
            - `headers` object: Headers to add to the request.
            - `query_params` object: Query parameters to append to the request URL.
            - `parameters` ToolParameter: The parameters the functions accepts, described as a JSON Schema object. See [JSON Schema reference](https://json-schema.org/understanding-json-schema/) for documentation about the format. Omitting parameters defines a function with an empty parameter list.
              - …
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: A mapping of variable names to JSON paths in the response body. These values will be extracted from the response and made available as dynamic variables for use.
            - `speak_during_execution` boolean: Determines whether the agent would say sentence like "One moment, let me check that." when executing the function. Recommend to turn on if your function call takes over 1s (including network) to complete, so that your agent remains responsive.
            - `speak_after_execution` boolean: Determines whether the agent would call LLM another time and speak when the result of function is obtained. Usually this needs to get turned on so user can get update for the function call.
            - `execution_message_description` string: The description for the sentence agent say during execution. Only applicable when speak_during_execution is true. Can write what to say or even provide examples. The default is "The message you will say to callee when calling this tool. Make sure it fits into the conversation smoothly.".
            - `execution_message_type` 'prompt' | 'static_text': Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `timeout_ms` integer: The maximum time in milliseconds the tool can run before it's considered timeout. If the tool times out, the agent would have that info. The minimum value allowed is 1000 ms (1 s), and maximum value allowed is 600,000 ms (10 min). By default, this is set to 120,000 ms (2 min).
            - `max_retry` integer: Maximum number of times to retry the request after a failed attempt, from 0 (no retry) to 5. Retries happen on any failure, with exponential backoff between attempts; the backoff delay is not configurable. `timeout_ms` applies per attempt rather than as a budget across all attempts, so an attempt that times out is still retried and the worst-case total duration is `timeout_ms` multiplied by (`max_retry` + 1) as well as any latency incurred by the exponential backoff + jitter between each retry. Only the final attempt's result is reported to the agent. Because retries repeat the request, only set this above 0 if your endpoint is idempotent — a retried request may be processed more than once. Defaults to 0 (no retry).
            - `args_at_root` boolean: If set to true, the parameters will be passed as root level JSON object instead of nested under "args".
            - `parameter_type` 'json' | 'form': How the tool's `parameters` are authored and shown in the dashboard editor — "form" for the visual parameter builder, "json" for a raw JSON Schema. Both produce the same `parameters` schema; this does not change how the request body is encoded (see `args_at_root`).
            - `enable_typing_sound` boolean: If true, play a typing sound on the agent audio track while this tool is executing. Useful when the tool takes a noticeable amount of time to prevent silence on the call.
            - `depends_on` string[]: Names of tools that must run before this one.
          - object: A pre-session tool (app / custom / code, discriminated by type) plus its dependency edges. The tool's name is the depends_on handle and must be unique across the agent's session tools. Parameters may set a constant value or a description for the LLM to infer the value. Pre-session inference uses the agent prompt and dynamic variables.
            - `type` 'code', required
            - `name` string, required: Name of the tool. Must be unique within all tools available to LLM at any given time (general tools + state tools + state edges). Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `description` string: Describes what this tool does and when to call this tool.
            - `code` string, required: JavaScript code to execute in the sandbox.
            - `timeout_ms` integer: The maximum time in milliseconds the code can run before it's considered timeout. Defaults to 30,000 ms (30 s).
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: A mapping of variable names to JSON paths in the code execution result. These mapped values will be extracted and added as dynamic variables.
            - `speak_during_execution` boolean: Determines whether the agent would say sentence like "One moment, let me check that." when executing the tool.
            - `speak_after_execution` boolean: Determines whether the agent would call LLM another time and speak when the result of function is obtained.
            - `execution_message_description` string: The description for the sentence agent say during execution. Only applicable when speak_during_execution is true.
            - `execution_message_type` 'prompt' | 'static_text': Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `enable_typing_sound` boolean: If true, play a typing sound on the agent audio track while this tool is executing.
            - `depends_on` string[]: Names of tools that must run before this one.
      - `post_session_tools` PostSessionTool[], nullable: Integration (Agent Functions) tools run as a dependency graph during teardown, after post-call analysis. Each tool can be gated by a condition. On calls the graph is stopped after five minutes so teardown can finish. Set to null to clear.
        - union
          - object: A post-session tool with an optional condition. Parameter inference uses the agent prompt, dynamic variables, and conversation. Same tools as PreSessionTool plus send_sms, which is only supported on phone calls — the session is over, so the speak settings are ignored and an inferred sms_content is written by the LLM from the transcript at teardown.
            - `type` 'integration_app', required
            - `name` string, required: Name of the tool. Must be unique within the phase's tools; referenced by depends_on. Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `app_id` string, required: The connection (App) this tool runs against. Must be a connection in the organization whose provider matches this tool's provider.
            - `provider` string, required: Provider of the connection. Must match the connection's provider; supported providers are listed by list-app-templates.
            - `app_tool_template_name` string, required: Name of the catalog template within the provider, as listed by list-app-templates.
            - `parameters` ToolParameter[]: The resolved input parameters, in order. Properties may pin a value with const (including {{variable}} references) or provide a description for LLM inference. Each property may also record selected_input_mode, the editor mode the user selected ("const_enum", "const_boolean", "const_value", "description_custom", or "description_preset"); it is stored and returned as-is, used only by the tool config UI. Omit the key when no mode is recorded; when set, const_* modes require a non-empty const, and description_* modes must omit const entirely. Each parameter's required list must match the schema returned by the corresponding step of the get-app-tool-schema loop.
              - …
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: Mapping of a dynamic-variable name to the response field (dot-path) it is populated from. Missing paths are ignored.
            - `output_selection` union: Which fields of a successful JSON tool response the agent and the transcript see. Omit to keep all fields. Does not affect response_variables, which are always extracted from the raw response. Non-JSON responses and failed tool results are not filtered. Response size limits still apply.
              - …
            - `description` string: Only applies to during conversation functions; ignored by the pre/post conversation. Overrides the catalog template's LLM-facing description.
            - `speak_during_execution` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. If true, will speak during execution.
            - `speak_after_execution` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. Determines whether the agent would call LLM another time and speak when the result of the tool is obtained.
            - `execution_message_description` string: Only applies to during conversation functions; ignored by the pre/post conversation. The message for the agent to speak when executing the tool. Only applicable when speak_during_execution is true.
            - `execution_message_type` 'prompt' | 'static_text': Only applies to during conversation functions; ignored by the pre/post conversation. Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `enable_typing_sound` boolean: Only applies to during conversation functions; ignored by the pre/post conversation. If true, play a typing sound on the agent audio track while this tool is executing. Useful when the tool takes a noticeable amount of time to prevent silence on the call.
            - `depends_on` string[]: Names of tools that must run before this one.
            - `condition` EquationCondition
              - …
          - object: A post-session tool with an optional condition. Parameter inference uses the agent prompt, dynamic variables, and conversation. Same tools as PreSessionTool plus send_sms, which is only supported on phone calls — the session is over, so the speak settings are ignored and an inferred sms_content is written by the LLM from the transcript at teardown.
            - `type` 'custom', required
            - `name` string, required: Name of the tool. Must be unique within all tools available to LLM at any given time (general tools + state tools + state edges). Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `url` string, required: Describes what the tool does, sometimes can also include information about when to call the tool.
            - `description` string: Describes what this tool does and when to call this tool.
            - `method` 'GET' | 'POST' | 'PUT' | 'PATCH' | 'DELETE': Method to use for the request, default to POST.
            - `headers` object: Headers to add to the request.
            - `query_params` object: Query parameters to append to the request URL.
            - `parameters` ToolParameter: The parameters the functions accepts, described as a JSON Schema object. See [JSON Schema reference](https://json-schema.org/understanding-json-schema/) for documentation about the format. Omitting parameters defines a function with an empty parameter list.
              - …
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: A mapping of variable names to JSON paths in the response body. These values will be extracted from the response and made available as dynamic variables for use.
            - `speak_during_execution` boolean: Determines whether the agent would say sentence like "One moment, let me check that." when executing the function. Recommend to turn on if your function call takes over 1s (including network) to complete, so that your agent remains responsive.
            - `speak_after_execution` boolean: Determines whether the agent would call LLM another time and speak when the result of function is obtained. Usually this needs to get turned on so user can get update for the function call.
            - `execution_message_description` string: The description for the sentence agent say during execution. Only applicable when speak_during_execution is true. Can write what to say or even provide examples. The default is "The message you will say to callee when calling this tool. Make sure it fits into the conversation smoothly.".
            - `execution_message_type` 'prompt' | 'static_text': Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `timeout_ms` integer: The maximum time in milliseconds the tool can run before it's considered timeout. If the tool times out, the agent would have that info. The minimum value allowed is 1000 ms (1 s), and maximum value allowed is 600,000 ms (10 min). By default, this is set to 120,000 ms (2 min).
            - `max_retry` integer: Maximum number of times to retry the request after a failed attempt, from 0 (no retry) to 5. Retries happen on any failure, with exponential backoff between attempts; the backoff delay is not configurable. `timeout_ms` applies per attempt rather than as a budget across all attempts, so an attempt that times out is still retried and the worst-case total duration is `timeout_ms` multiplied by (`max_retry` + 1) as well as any latency incurred by the exponential backoff + jitter between each retry. Only the final attempt's result is reported to the agent. Because retries repeat the request, only set this above 0 if your endpoint is idempotent — a retried request may be processed more than once. Defaults to 0 (no retry).
            - `args_at_root` boolean: If set to true, the parameters will be passed as root level JSON object instead of nested under "args".
            - `parameter_type` 'json' | 'form': How the tool's `parameters` are authored and shown in the dashboard editor — "form" for the visual parameter builder, "json" for a raw JSON Schema. Both produce the same `parameters` schema; this does not change how the request body is encoded (see `args_at_root`).
            - `enable_typing_sound` boolean: If true, play a typing sound on the agent audio track while this tool is executing. Useful when the tool takes a noticeable amount of time to prevent silence on the call.
            - `depends_on` string[]: Names of tools that must run before this one.
            - `condition` EquationCondition
              - …
          - object: A post-session tool with an optional condition. Parameter inference uses the agent prompt, dynamic variables, and conversation. Same tools as PreSessionTool plus send_sms, which is only supported on phone calls — the session is over, so the speak settings are ignored and an inferred sms_content is written by the LLM from the transcript at teardown.
            - `type` 'code', required
            - `name` string, required: Name of the tool. Must be unique within all tools available to LLM at any given time (general tools + state tools + state edges). Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `description` string: Describes what this tool does and when to call this tool.
            - `code` string, required: JavaScript code to execute in the sandbox.
            - `timeout_ms` integer: The maximum time in milliseconds the code can run before it's considered timeout. Defaults to 30,000 ms (30 s).
            - `execution_status_variable` string: Dynamic variable to store execution status as "success" or "failure". Tools skipped by conditions or dependencies leave it unchanged.
            - `response_variables` object: A mapping of variable names to JSON paths in the code execution result. These mapped values will be extracted and added as dynamic variables.
            - `speak_during_execution` boolean: Determines whether the agent would say sentence like "One moment, let me check that." when executing the tool.
            - `speak_after_execution` boolean: Determines whether the agent would call LLM another time and speak when the result of function is obtained.
            - `execution_message_description` string: The description for the sentence agent say during execution. Only applicable when speak_during_execution is true.
            - `execution_message_type` 'prompt' | 'static_text': Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `enable_typing_sound` boolean: If true, play a typing sound on the agent audio track while this tool is executing.
            - `depends_on` string[]: Names of tools that must run before this one.
            - `condition` EquationCondition
              - …
          - object: A post-session tool with an optional condition. Parameter inference uses the agent prompt, dynamic variables, and conversation. Same tools as PreSessionTool plus send_sms, which is only supported on phone calls — the session is over, so the speak settings are ignored and an inferred sms_content is written by the LLM from the transcript at teardown.
            - `name` string, required: Name of the tool. Must be unique within all tools available to LLM at any given time (general tools + state tools + state edges). Must be consisted of a-z, A-Z, 0-9, or contain underscores and dashes, with a maximum length of 64 (no space allowed).
            - `type` 'send_sms', required
            - `description` string: Describes what the tool does, sometimes can also include information about when to call the tool.
            - `speak_during_execution` boolean: If true, the agent will speak a short line before sending the SMS. If omitted, defaults to true (same as end_call / transfer_call tools).
            - `execution_message_description` string: Describes what to say before sending the SMS. Only applicable when speak_during_execution is true.
            - `execution_message_type` 'prompt' | 'static_text': Type of execution message. "prompt" means the agent will use execution_message_description as a prompt to generate the message. "static_text" means the agent will speak the execution_message_description directly. Defaults to "prompt".
            - `sms_content` union, required
              - …
            - `sms_sender` union: SMS sender number configuration. If omitted, the message is sent from the agent's phone number.
              - …
            - `sms_destination` union: SMS destination number configuration. If omitted, the message is sent to the customer number on the call.
              - …
            - `speak_after_execution` boolean: Whether the agent should speak after the SMS tool finishes execution.
            - `depends_on` string[]: Names of tools that must run before this one.
            - `condition` EquationCondition
              - …
    - `retell_llm` RetellLlmOverride: Override properties for Retell LLM configuration in agent override requests.
      - `model_config` union: Model settings for the response engine. type decides which models primary and backup accept. On update, each top-level key replaces the current one and omitted keys are kept; set backup to null to remove it. Changing type requires primary and clears backup, while temperature carries over. In agent_override, type is required whenever primary is set; setting backup to null disables it for that call only.
        - CascadingModelConfig
          - `type` 'cascading': Runs a text LLM with a separate voice. This is the default type.
          - `primary` TextModelRef
            - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
            - `tier` 'standard' | 'fast' | 'ultrafast': Capacity tier for a model. standard is always available. fast draws from dedicated capacity for lower, steadier latency, and ultrafast is faster still on voice agents; faster tiers cost more. Each model offers its own set of tiers, and a tier the model does not offer is rejected.
          - `backup` union
            - BackupModelRef: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
              - …
            - unknown
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
        - S2SModelConfig
          - `type` 's2s': Runs a speech-to-speech model.
          - `primary` S2SModelRef
            - `model` 'gpt-realtime-2.1' | 'gpt-realtime-2.1-mini' | 'gpt-realtime-2' | 'gpt-realtime-1.5' | 'gpt-realtime' | 'gpt-realtime-mini' | 'grok-voice-3.0', required: The speech-to-speech model to use
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
        - DuplexModelConfig: Retell LLMs only. A full-duplex voice model talks to the user and delegates to a cascading backend model, which runs the LLM's prompt, tools and knowledge base. Requires backend, and the Retell LLM's voice_prompt. Tools' speak_during_execution is ignored, since the voice model covers the wait itself, and speak_after_execution is always on.
          - `type` 'duplex': Runs a full-duplex voice model with a cascading backend.
          - `primary` DuplexModelRef
            - `model` 'gpt-live-1', required
          - `backend` TextModelRef
            - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
            - `tier` 'standard' | 'fast' | 'ultrafast': Capacity tier for a model. standard is always available. fast draws from dedicated capacity for lower, steadier latency, and ultrafast is faster still on voice agents; faster tiers cost more. Each model offers its own set of tiers, and a tier the model does not offer is rejected.
          - `backup` union
            - BackupModelRef: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
              - …
            - unknown
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
      - `backup_model` union
        - BackupModelChoice: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
          - `type` 'cascading', required: Type of model choice
          - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
          - `high_priority` boolean: Whether to use high priority pool with more dedicated resource, default false
          - `ultrafast` boolean: Whether to use the ultrafast tier for much faster response generation, default false. Only supported by gpt-6.1-sol on voice agents; ignored otherwise. Takes precedence over high_priority.
          - `backup_kick_in_timeout` integer: Duration in milliseconds to prefer primary model output while both models run concurrently. Must be between 300 and 60000 milliseconds, inclusive. Defaults to 1500 when omitted.
        - unknown
      - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon' | 'null', nullable: Available LLM models for agents.
      - `s2s_model` 'gpt-realtime-2.1' | 'gpt-realtime-2.1-mini' | 'gpt-realtime-2' | 'gpt-realtime-1.5' | 'gpt-realtime' | 'gpt-realtime-mini' | 'grok-voice-3.0' | 'null', nullable: The speech-to-speech model to use
      - `model_temperature` number: Deprecated, use model_config.temperature. Ignored when model_config is set.
      - `model_high_priority` boolean, nullable: Deprecated, use model_config.primary.tier fast. Ignored when model_config is set.
      - `model_ultrafast` boolean, nullable: Deprecated, use model_config.primary.tier ultrafast. Ignored when model_config is set.
      - `knowledge_base_ids` string[], nullable: A list of knowledge base ids to use for this resource.
      - `kb_config` KBConfig
        - `top_k` integer: Max number of knowledge base chunks to retrieve
        - `filter_score` number: Similarity threshold for filtering search results
      - `start_speaker` 'user' | 'agent': The speaker who starts the conversation. Required. Must be either 'user' or 'agent'.
      - `begin_after_user_silence_ms` integer, nullable: If set, the AI will begin the conversation after waiting for the user for the duration (in milliseconds) specified by this attribute. This only applies if the agent is configured to wait for the user to speak first. If not set, the agent will wait indefinitely for the user to speak.
      - `begin_message` string, nullable: First utterance said by the agent in the call. If not set, LLM will dynamically generate a message. If set to "", agent will wait for user to speak first.
    - `conversation_flow` ConversationFlowOverride: Override properties for conversation flow configuration in agent override requests.
      - `model_config` union: Model settings for the response engine. type decides which models primary and backup accept. On update, each top-level key replaces the current one and omitted keys are kept; set backup to null to remove it. Changing type requires primary and clears backup, while temperature carries over. In agent_override, type is required whenever primary is set; setting backup to null disables it for that call only.
        - CascadingModelConfig
          - `type` 'cascading': Runs a text LLM with a separate voice. This is the default type.
          - `primary` TextModelRef
            - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
            - `tier` 'standard' | 'fast' | 'ultrafast': Capacity tier for a model. standard is always available. fast draws from dedicated capacity for lower, steadier latency, and ultrafast is faster still on voice agents; faster tiers cost more. Each model offers its own set of tiers, and a tier the model does not offer is rejected.
          - `backup` union
            - BackupModelRef: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
              - …
            - unknown
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
        - S2SModelConfig
          - `type` 's2s': Runs a speech-to-speech model.
          - `primary` S2SModelRef
            - `model` 'gpt-realtime-2.1' | 'gpt-realtime-2.1-mini' | 'gpt-realtime-2' | 'gpt-realtime-1.5' | 'gpt-realtime' | 'gpt-realtime-mini' | 'grok-voice-3.0', required: The speech-to-speech model to use
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
        - DuplexModelConfig: Retell LLMs only. A full-duplex voice model talks to the user and delegates to a cascading backend model, which runs the LLM's prompt, tools and knowledge base. Requires backend, and the Retell LLM's voice_prompt. Tools' speak_during_execution is ignored, since the voice model covers the wait itself, and speak_after_execution is always on.
          - `type` 'duplex': Runs a full-duplex voice model with a cascading backend.
          - `primary` DuplexModelRef
            - `model` 'gpt-live-1', required
          - `backend` TextModelRef
            - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
            - `tier` 'standard' | 'fast' | 'ultrafast': Capacity tier for a model. standard is always available. fast draws from dedicated capacity for lower, steadier latency, and ultrafast is faster still on voice agents; faster tiers cost more. Each model offers its own set of tiers, and a tier the model does not offer is rejected.
          - `backup` union
            - BackupModelRef: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
              - …
            - unknown
          - `temperature` number: Randomness of the response, from 0 to 1. Defaults to 0; a lower value is recommended for tool calling.
      - `backup_model` union
        - BackupModelChoice: Backup models are supported only for call agents. Response engines with a backup model cannot be used by chat agents.
          - `type` 'cascading', required: Type of model choice
          - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
          - `high_priority` boolean: Whether to use high priority pool with more dedicated resource, default false
          - `ultrafast` boolean: Whether to use the ultrafast tier for much faster response generation, default false. Only supported by gpt-6.1-sol on voice agents; ignored otherwise. Takes precedence over high_priority.
          - `backup_kick_in_timeout` integer: Duration in milliseconds to prefer primary model output while both models run concurrently. Must be between 300 and 60000 milliseconds, inclusive. Defaults to 1500 when omitted.
        - unknown
      - `model_choice` union
        - ModelChoiceCascading
          - `type` 'cascading', required: Type of model choice
          - `model` 'gpt-4.1' | 'gpt-4.1-mini' | 'gpt-5' | 'gpt-5-mini' | 'gpt-5-nano' | 'gpt-5.1' | 'gpt-5.2' | 'gpt-5.4' | 'gpt-5.4-mini' | 'gpt-5.4-nano' | 'gpt-5.5' | 'gpt-5.6-terra' | 'gpt-5.6-luna' | 'gpt-6-astra' | 'gpt-6-sol' | 'gpt-6.1-sol' | 'gpt-6.1-terra' | 'gpt-6-luna' | 'claude-4.6-sonnet' | 'claude-5-opus' | 'claude-5.5-opus' | 'claude-5-sonnet' | 'claude-5.5-sonnet' | 'claude-5.5-haiku' | 'claude-4.5-haiku' | 'gemini-3.0-flash' | 'gemini-3.1-flash-lite' | 'gemini-3.5-flash' | 'gemini-3.5-flash-lite' | 'gemini-3.6-flash' | 'gemini-3.7-flash' | 'gemini-3.8-flash' | 'gemini-4-argon', required: Available LLM models for agents.
          - `high_priority` boolean: Whether to use high priority pool with more dedicated resource, default false
          - `ultrafast` boolean: Whether to use the ultrafast tier for much faster response generation, default false. Only supported by gpt-6.1-sol on voice agents; ignored otherwise. Takes precedence over high_priority.
        - ModelChoiceS2S
          - `type` 's2s', required: Type of model choice
          - `model` 'gpt-realtime-2.1' | 'gpt-realtime-2.1-mini' | 'gpt-realtime-2' | 'gpt-realtime-1.5' | 'gpt-realtime' | 'gpt-realtime-mini' | 'grok-voice-3.0', required: The speech-to-speech model to use
      - `model_temperature` number, nullable: Deprecated, use model_config.temperature. Ignored when model_config is set.
      - `knowledge_base_ids` string[], nullable: Knowledge base IDs for RAG (Retrieval-Augmented Generation).
      - `kb_config` KBConfig
        - `top_k` integer: Max number of knowledge base chunks to retrieve
        - `filter_score` number: Similarity threshold for filtering search results
      - `start_speaker` 'user' | 'agent': Who starts the conversation - user or agent.
      - `begin_after_user_silence_ms` integer, nullable: If set, the AI will begin the conversation after waiting for the user for the duration (in milliseconds) specified by this attribute. This only applies if the agent is configured to wait for the user to speak first. If not set, the agent will wait indefinitely for the user to speak.
  - `metadata` object: An arbitrary object for storage purpose only. You can put anything here like your internal customer id associated with the call. Not used for processing. You can later get this field from the call object.
  - `retell_llm_dynamic_variables` object: Add optional dynamic variables in key value pairs of string that injects into your Response Engine prompt and tool description. Only applicable for Response Engine.
  - `current_node_id` string, nullable: Start the call at this conversation flow node (stage). Must be a valid node id in the agent's conversation flow. Only applicable when the agent uses conversation flow as the response engine. Ignored for retell-llm agents.
  - `current_state` string, nullable: Start the conversation in this state (stage). Must be a valid state name in the agent's Retell LLM. Only applicable when the agent uses Retell LLM with states. Ignored for conversation-flow agents.

## Response `201`

Successfully created a web call.

- V3CreateWebCallResponse
  - `call_id` string, required: Unique identifier for the web call.
  - `access_token` string, required: Token authorizing this browser to join the web call. Pass it to your frontend.
  - `transport` 'gateway', required: Connection transport to select in the web client.
  - `ice_servers` IceServer[], required: ICE servers to configure before the browser creates its peer connection.
    - `urls` union, required
      - string
      - string[]
    - `username` string
    - `credential` string
  - `expires_at` integer, required: Unix epoch ms when the access_token expires.

## Other responses

- `400`: Bad Request
- `401`: Unauthorized
- `402`: Payment Required
- `422`: Unprocessable Content
- `429`: Too Many Requests
- `500`: Internal Server Error

## Changes

- **2026-10-31** `de3a1d542da5`: 2 warning
  - removed the request property `agent_override/conversation_flow/model_config/anyOf[#/components/schemas/DuplexModelConfig]/voice_prompt`
  - removed the request property `agent_override/retell_llm/model_config/anyOf[#/components/schemas/DuplexModelConfig]/voice_prompt`
- **2026-10-30** `5d018ad1bc4d`: 2 info
  - added `#/components/schemas/DuplexModelConfig` to the `agent_override/conversation_flow/model_config` request property `anyOf` list
  - added `#/components/schemas/DuplexModelConfig` to the `agent_override/retell_llm/model_config` request property `anyOf` list
- **2026-10-29** `b982b63b07d0`: 2 warning, 1 info
  - removed the request property `agent_override/conversation_flow/tool_call_strict_mode`
  - removed the request property `agent_override/retell_llm/tool_call_strict_mode`
  - added the new `drama-3` enum value to the request property `agent_override/agent/voice_model`
- …earlier changes not shown

[Operation history](https://skmtc.dev/retellai/apis/retell-sdk/changes/v3/create-web-call/post.md)

---

[API](https://skmtc.dev/retellai/apis/retell-sdk.md) - [All operations](https://skmtc.dev/retellai/apis/retell-sdk/llms.txt) - [OpenAPI document](https://skmtc.dev/retellai/apis/retell-sdk/revisions/e482acaa292c?raw)