Call Data API

The Call Data API reads the results of your voice calls — every call a published voice agent handled, whether it came from a campaign, an outbound dispatch, an inbound number, or the web widget. It is the programmatic equivalent of the Call Logs screen and the call detail it opens: the same rows, the same tabs, the same numbers.

Every endpoint is a GET. Nothing here creates, changes, or deletes anything, so every request is safe to retry.

When to Use This

Pull call outcomes into your CRM, data warehouse, or BI tool instead of exporting by hand.
Fetch a transcript or a recording for a call your own system knows the ID of.
Read what the agent extracted (structured output) and act on it downstream.
Audit cost, latency, and the event timeline of a call from your own tooling.
Build a campaign report that includes contacts that were never dialed, and why.

Which endpoint answers which question?

You want	Use
Every call in a date range, with status, duration, and credits	List calls
Everything about one call in a single request	Get a call
Just the conversation, as text or turns	Get transcript
What the agent extracted at the end of the call	Get structured output
The audio file	Download recording
Every dial attempt of a campaign, including contacts that were never called	List attempts on the Campaign API

Prerequisites

An API Key with the Voice Agent scope. If the key uses a specific Agent Access list, only calls handled by agents on that list are visible — the list is narrowed to them automatically, and a per-call request for any other agent's call fails with 403.
A call_id, for the per-call endpoints. It comes from List calls, from List attempts for a campaign, or from the call_data.call_id field of the Post-Webhook payload.

Authentication

Every request must include the api-key header — see API Keys for how to generate and manage tokens.

api-key: sk_<your-secret>

Every call is recorded in Request Logs, including the failures, with a matching error code.

Base path

Call endpoints are served at the root of your DronaHQ host — https://<your-dronahq-host>/voice/calls. There is no /api/v1 prefix.

This is a server-to-server API

Browsers cannot call these endpoints directly: the api-key header is not on the CORS allow-list, so a request from browser JavaScript fails its preflight. Call the API from your backend and never ship an API key to a browser or mobile client — a key is a workspace-wide secret. This applies to the recording endpoint too: proxy the audio through your own server rather than pointing an <audio> tag at it.

Endpoints

Method	Path	What it returns
`GET`	`/voice/calls`	List calls — the Call Logs table
`GET`	`/voice/calls/{call_id}`	Get a call — the whole call detail in one response
`GET`	`/voice/calls/{call_id}/transcript`	Get transcript — the Transcripts tab
`GET`	`/voice/calls/{call_id}/events`	Get events — the Logs tab
`GET`	`/voice/calls/{call_id}/structured_output`	Get structured output — the Structured Outputs tab
`GET`	`/voice/calls/{call_id}/messages`	Get messages — the Messages tab
`GET`	`/voice/calls/{call_id}/latency`	Get latency — the Latency Summary tab
`GET`	`/voice/calls/{call_id}/cost`	Get cost — the Call Cost tab
`GET`	`/voice/calls/{call_id}/recording`	Download recording — the audio player

The per-call endpoints map one-to-one onto the tabs of the call detail in the product, and Get a call returns all of them together. Both are produced by the same code, so a field has the same shape whichever way you fetch it.

For campaigns there is one more read endpoint, documented with the rest of the Campaign API: GET /voice/campaigns/{campaign_uuid}/attempts, the campaign's own dial registry.

Retention

Call data is kept for 30 days

Every endpoint on this page reads from call telemetry that expires 30 days after the call started. After that the call is gone, not hidden: it drops out of the list and every per-call endpoint returns 404, including the recording.

Pull what you need into your own storage within that window. If you run campaigns, the attempts endpoint is permanent and still lists the attempt and its outcome after the 30 days, without the transcript, timings, or recording.

Pagination

The two list-shaped responses — List calls and Get events — page with limit and offset and report has_more. There is no total.

Page on has_more, never on page length

Keep requesting with offset += limit while has_more is true. Do not stop because a page came back shorter than limit.

List calls

GET /voice/calls

Every call the key can see, newest first — with the same columns as the Call Logs table, including the credits breakdown.

Query parameters

All optional. Parameters marked repeatable can be given more than once (?type=web&type=inbound-phone) and are ORed.

Parameter	Type	Default	Description
`agent_uuid`	string, repeatable	—	Only calls handled by these agents. A key with an Agent Access list may only name agents on it (`403` otherwise)
`campaign_uuid`	string	—	Only calls placed by this campaign
`start_datetime`	ISO-8601 datetime	—	Calls that started at or after this instant
`end_datetime`	ISO-8601 datetime	—	Calls that started at or before this instant
`status`	string	—	`completed` or `failed`
`type`	string, repeatable	—	`web`, `inbound-phone`, `outbound-phone`
`direction`	string, repeatable	—	`inbound`, `outbound`
`ended_reason`	string, repeatable	—	Exact match on the end reason
`hangup_cause_code`	string, repeatable	—	The carrier's hangup code, e.g. `4000`
`hangup_source`	string, repeatable	—	Who hung up per the carrier, e.g. `Callee`
`search`	string	—	Substring match on `call_id`
`sort_key`	string	`start_time`	`start_time`, `duration_ms`, or `cost`
`sort_dir`	string	`desc`	`asc` or `desc`
`limit`	integer	`25`	1–500
`offset`	integer	`0`	For paging

Always send a date range

Both bounds are optional — without them the first page is simply the newest calls — but a bounded window is what makes a call-log query cheap and what a report actually wants. Timestamps without an offset are read as UTC; 2026-09-01T00:00:00+05:30 is honored as-is. A start_datetime later than end_datetime is rejected with 422.

Unrecognized filter values return nothing, not an error

status accepts only completed and failed; anything else is ignored. The repeatable filters are exact matches, so a typo in type or ended_reason returns an empty page rather than a 400.

Example request

curl -X GET 'https://<your-dronahq-host>/voice/calls?start_datetime=2026-09-01T00:00:00Z&end_datetime=2026-09-30T23:59:59Z&limit=25&offset=0' \
  -H 'api-key: sk_your-generated-secret-here'

Only one campaign's calls, failed ones first:

curl -X GET 'https://<your-dronahq-host>/voice/calls?campaign_uuid=<campaign-uuid>&status=failed&limit=100' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_logs": [
    {
      "id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
      "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
      "agent_uuid": "a1d3...",
      "agent_name": null,
      "campaign_uuid": "0f2c9c1e-7a3d-4f80-9c11-2b6d5a7e91aa",
      "has_recording": true,
      "tools_used": 1,
      "user_number": "+919876543210",
      "from_number": "+918041234567",
      "to_number": "+919876543210",
      "type": "outbound-phone",
      "direction": "outbound",
      "end_reason": "customer-ended-call",
      "end_reason_detail": "",
      "hangup_cause_code": "4000",
      "hangup_cause_name": "Normal Hangup",
      "hangup_source": "Callee",
      "stir_verification": "",
      "answered_by": "human",
      "status": "completed",
      "start_time": "2026-09-16T04:12:09.120000+00:00",
      "end_time": "2026-09-16T04:13:01.004000+00:00",
      "duration_ms": 51884,
      "cost": 0.0,
      "credits": 1.42,
      "llm_credits": 0.91,
      "stt_credits": 0.27,
      "tts_credits": 0.24,
      "structured_output_credits": 0.0
    }
  ],
  "has_more": true
}

The call object

Every row of the list, and the base of Get a call.

Field	Type	Description
`call_id`	string	The call's identifier. `id` is the same value
`agent_uuid`	string	The voice agent that handled the call
`agent_name`	null	Always `null` on this API — look the agent up by `agent_uuid`
`campaign_uuid`	string or null	The campaign that placed the call; `null` for dispatch, inbound, and web calls
`has_recording`	boolean	Whether Download recording has something to return
`tools_used`	integer	How many tool calls the agent made
`user_number`	string or null	The customer's number — `to_number` on an outbound call, `from_number` on an inbound one. `null` for web calls
`from_number` / `to_number`	string or null	The two legs as dialed
`type`	string	`web`, `inbound-phone`, or `outbound-phone`
`direction`	string	`inbound` or `outbound`
`end_reason`	string	Why the call ended — see End reasons
`end_reason_detail`	string	The provider's own words when a dial failed; empty otherwise
`hangup_cause_code` / `hangup_cause_name` / `hangup_source`	string	The carrier's disconnect metadata. Empty when the provider sent none
`stir_verification`	string	STIR/SHAKEN attestation as reported by Twilio; empty on other providers
`answered_by`	string	Answering-machine detection on an outbound dial: `human`, `machine`, or empty when not classified
`status`	string	`completed` or `failed`
`start_time` / `end_time`	ISO-8601 or null	UTC
`duration_ms`	integer or null	Talk time in milliseconds; `0` or `null` for a call that never connected
`credits`	number or null	Total credits billed for the call
`llm_credits` / `stt_credits` / `tts_credits` / `structured_output_credits`	number or null	The split of `credits` by cost center — the same four the Call Cost tab shows
`cost`	number	Legacy field, always `0.0`. Use `credits`

Credits can be null

Credits are looked up from the billing system per call when you request the list. When that lookup is unavailable the four credit fields are null — never a fake 0. A call that never engaged the agent (no-answer, busy, failed-dial, …) is a genuine 0.

End reasons

end_reason is a closed vocabulary set by the platform, so you can filter and group on it:

Value	Meaning
`customer-ended-call`	The caller hung up
`agent-ended-call`	The agent ended the call after completing its objective
`silence-timeout`	Nobody spoke for the configured silence window
`max-duration-reached`	The agent's maximum call duration was hit
`voicemail-detected` / `amd-machine-detected`	Voicemail or an answering machine picked up
`transferred` / `transfer-completed`	The call was handed to a human
`no-answer` / `busy` / `declined` / `failed-dial` / `carrier-hangup`	The dial did not connect, or the carrier ended it — `end_reason_detail` and the `hangup_cause_*` fields carry the carrier's reason
`agent-error` / `transcriber-error` / `llm-error` / `tts-error` / `transport-error` / `worker-shutdown`	The platform ended the call

Twilio numbers report SIP phrases for unconnected dials

On calls placed through Twilio, the coarse no-answer / busy / failed-dial / carrier-hangup buckets are replaced by the SIP response phrase, slugified — a 480 is temporarily-unavailable, a 486 is busy-here. Filter on the values you actually see in your own data rather than assuming the list above is exhaustive for Twilio.

Get a call

GET /voice/calls/{call_id}

One call in full, in a single request: the call object plus everything the per-call endpoints below return separately.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call": {
    "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
    "agent_uuid": "a1d3...",
    "campaign_uuid": "0f2c9c1e-7a3d-4f80-9c11-2b6d5a7e91aa",
    "status": "completed",
    "end_reason": "customer-ended-call",
    "duration_ms": 51884,
    "credits": 1.42,

    "provider": "plivo",
    "agent_version_id": "v7",
    "transcript_text": "Agent: Hello Rahul, I'm calling about your pending payment of 1500 rupees.\nUser: Yes, I'll pay today.\nAgent: Thank you. I've noted that. Goodbye!\n",
    "messages": [
      { "role": "agent", "content": "Hello Rahul, I'm calling about your pending payment of 1500 rupees.", "time_ms": 1900, "seconds_from_start": 1.9, "duration_ms": 4100, "tool_calls": [] },
      { "role": "user", "content": "Yes, I'll pay today.", "time_ms": 7200, "seconds_from_start": 7.2, "duration_ms": 1300, "tool_calls": [] }
    ],
    "structured_output": [
      { "id": "so_payment_confirmed", "name": "payment_confirmed", "schema_type": "boolean", "status": "success", "result": true, "reason": null }
    ],
    "latency": {
      "turnLatencyAverage": 812,
      "transcriberLatencyAverage": 210,
      "modelLatencyAverage": 380,
      "voiceLatencyAverage": 222,
      "turnLatencies": [
        { "turnId": "t1", "transcriberLatency": 205, "endpointingLatency": 188, "modelFirstTokenLatency": 240, "modelTotalLatency": 371, "voiceFirstAudioLatency": 219, "turnLatency": 795 }
      ]
    },
    "consumption": {
      "llm": { "input_tokens": 3120, "cached_input_tokens": 0, "output_tokens": 214, "model": "gpt-4o-mini", "provider": "openai" },
      "stt": { "duration_seconds": 51.9, "model": "nova-3", "provider": "deepgram" },
      "tts": { "character_count": 148, "model": "eleven_flash_v2_5", "provider": "elevenlabs" },
      "structured_output": { "input_tokens": 412, "output_tokens": 6, "model": "gpt-4o-mini", "provider": "openai" }
    },
    "consumption_detail": {
      "billed": { "llm": { "input_tokens": 3120, "output_tokens": 214, "provider": "openai", "model": "gpt-4o-mini" } },
      "unbilled": { "skills": {}, "customer_key": {} },
      "tts_verification": {}
    },
    "variables": {
      "customer_name": "Rahul",
      "outstanding_amount": "1500",
      "dispatch": {
        "campaign_uuid": "0f2c9c1e-7a3d-4f80-9c11-2b6d5a7e91aa",
        "campaign_row_index": 12,
        "campaign_fields": { "number": "+919876543210", "customer_name": "Rahul", "outstanding_amount": "1500" },
        "is_alternate_number": false
      }
    },
    "campaign_fields": { "number": "+919876543210", "customer_name": "Rahul", "outstanding_amount": "1500" }
  }
}

Shortened: every field of the call object is present, and each nested payload is the complete object its own endpoint returns — see the sections below for the full shapes.

What it adds to the call object

Field	Same as	Description
`provider`	—	The telephony provider that carried the call (`plivo`, `twilio`, …); empty for web calls
`agent_version_id`	—	The published version of the agent that took the call
`transcript_text`	`/transcript?format=text`	The rendered `User: … / Agent: …` transcript
`messages`	`/messages`	The raw per-turn message log
`structured_output`	`/structured_output`	The extraction results
`latency`	`/latency`	Per-turn latencies and averages
`consumption`, `consumption_detail`	`/cost`	Raw usage and its billed / unbilled split. Credits are on the call object itself
`variables`	—	The variable map the agent ran with — your dynamic variables, plus a `dispatch` block on campaign and dispatch calls
`campaign_fields`	—	For a campaign call, the contact's CSV row, lifted out of `variables.dispatch` so you need not know the nesting. `null` otherwise

Two things are not in this response

Events — the timeline can run to hundreds of rows, so it is only available paged from Get events. The recording — audio, from Download recording. has_recording tells you whether there is one.

Get transcript

GET /voice/calls/{call_id}/transcript

The conversation, in one of three shapes. Pick with format.

`format`	Returns	Use it when
`text` (default)	`transcript` — one string, `User: …` / `Agent: …` lines	You want to display or store the transcript as-is
`turns`	`turns` — the spoken lines as `[{role, text, seconds_from_start, duration_ms}]`	You want to process it turn by turn
`full`	`messages` — the raw per-turn log, identical to Get messages	You need tool calls, word confidences, or unspoken model output

Only the key for the requested format is present in the response; the other two are omitted, not null.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/transcript?format=turns' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "format": "turns",
  "turns": [
    { "role": "agent", "text": "Hello Rahul, I'm calling about your pending payment of 1500 rupees.", "seconds_from_start": 1.9, "duration_ms": 4100 },
    { "role": "user", "text": "Yes, I'll pay today.", "seconds_from_start": 7.2, "duration_ms": 1300 },
    { "role": "agent", "text": "Thank you. I've noted that. Goodbye!", "seconds_from_start": 9.4, "duration_ms": 2200 }
  ]
}

turns contains only what was actually said on the call — role is user or agent — ordered by seconds_from_start. A call with no conversation (never connected, or the agent has transcripts disabled) still returns 200, with an empty transcript, turns, or messages. An unknown format is a 422.

Get events

GET /voice/calls/{call_id}/events

The call's event timeline — what the Logs tab shows. Every pipeline, telephony, transcriber, model, voice, tool, webhook, and skill event in order, each with the turn it belongs to and structured attributes.

Query parameters

Parameter	Type	Default	Description
`limit`	integer	`1000`	1–5000
`offset`	integer	`0`	For paging

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/events?limit=1000&offset=0' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "events": [
    {
      "timestamp": "2026-09-16T04:12:16.902000+00:00",
      "seq": 41,
      "category": "transcriber",
      "event_name": "stt.final_transcript",
      "turn_id": "t2",
      "span_id": "sp-41",
      "parent_span_id": null,
      "severity": "info",
      "body": "Yes, I'll pay today.",
      "attributes": { "language": "en", "confidence": 0.97, "latency_ms": 205 },
      "consumption": null,
      "skill_selection": null,
      "skill_execution": null,
      "skill_extraction": null
    },
    {
      "timestamp": "2026-09-16T04:12:17.288000+00:00",
      "seq": 42,
      "category": "model",
      "event_name": "llm.generation",
      "turn_id": "t2",
      "span_id": "sp-42",
      "parent_span_id": "sp-41",
      "severity": "info",
      "body": "Thank you. I've noted that. Goodbye!",
      "attributes": { "model": "gpt-4o-mini", "first_token_ms": 240, "total_ms": 371 },
      "consumption": { "input_tokens": 1180, "output_tokens": 12 },
      "skill_selection": null,
      "skill_execution": null,
      "skill_extraction": null
    }
  ],
  "has_more": false
}

Field	Description
`seq`	Position in the timeline — sort on this, not on `timestamp`
`category`	`call`, `telephony`, `pipeline`, `transcriber`, `model`, `voice`, `knowledge_base`, `memory`, `latency`, `transport`, `structured_output`, `webhook`, `skill`, `system`
`event_name`	The specific event within the category, e.g. `stt.final_transcript`
`turn_id`	Groups the events of one conversational turn; `null` for call-level events
`span_id` / `parent_span_id`	Trace nesting, where the event is part of a span
`severity`	`debug`, `info`, `warn`, `error`, `fatal`
`body`	A human-readable line, where the event has one
`attributes`	Structured detail, specific to the event
`consumption`	Tokens or units consumed by this event, where a model ran
`skill_selection` / `skill_execution` / `skill_extraction`	Present on skill events only

A typical call has hundreds of events

Page with offset / limit and has_more. The default limit of 1000 covers most calls in one request.

Get structured output

GET /voice/calls/{call_id}/structured_output

The end-of-call extraction results — what the Structured Outputs tab shows. One entry per structured output configured on the agent. An agent with none configured returns an empty list.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/structured_output' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "structured_output": [
    {
      "id": "so_payment_confirmed",
      "name": "payment_confirmed",
      "input_type": "ai",
      "schema_type": "boolean",
      "status": "success",
      "result": true,
      "reason": null,
      "model": "gpt-4o-mini",
      "provider": "openai",
      "metrics": { "input_tokens": 412, "output_tokens": 6, "total_tokens": 418 },
      "last_committed_at": "2026-09-16T04:13:03Z"
    }
  ]
}

Field	Description
`id` / `name`	The structured output as configured on the agent
`schema_type`	The type `result` is cast to — `boolean`, `string`, `enum`, `number`, `integer`, `array`, or `object`
`status`	`success`, `skipped`, `failed`, or `pending` — see below
`result`	The extracted value. Only meaningful when `status` is `success`
`reason`	Why extraction failed or was skipped
`model` / `provider` / `metrics`	What ran the extraction and what it consumed

Read status before result

`status`	Meaning
`success`	`result` holds the extracted value, typed per `schema_type`
`skipped`	There was no conversation to extract from — `result` is empty
`failed`	Extraction ran and failed; `reason` says why
`pending`	The extraction step never got to mark this output — treat as unknown

result on anything but success is not a value your system should act on.

Get messages

GET /voice/calls/{call_id}/messages

The raw per-turn message log — what the Messages tab shows. Unlike the transcript this includes every role: the system prompt, what the model generated but never said aloud, tool calls and their results, with timing, word-level confidences, and the skill that was driving each turn. The same array Get transcript returns for format=full.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/messages' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "messages": [
    {
      "role": "agent",
      "content": "Hello Rahul, I'm calling about your pending payment of 1500 rupees.",
      "time_ms": 1900,
      "seconds_from_start": 1.9,
      "duration_ms": 4100,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "Yes, I'll pay today.",
      "time_ms": 7200,
      "seconds_from_start": 7.2,
      "duration_ms": 1300,
      "tool_calls": []
    }
  ]
}

`role`	Meaning
`system`	The system prompt the agent ran with
`user`	Something the caller said
`agent`	Something the agent said aloud
`agent-unspoken`	A reply the model generated (and was billed for) that the caller never heard — superseded by an interruption or a re-fire
`tool-call` / `tool-result`	A tool the agent invoked, and what came back

Only user and agent make it into the transcript. Use this endpoint when you need the rest.

Get latency

GET /voice/calls/{call_id}/latency

Per-turn latency and the call's averages, in milliseconds — what the Latency Summary tab shows.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/latency' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "latency": {
    "turnLatencyAverage": 812,
    "transcriberLatencyAverage": 210,
    "modelLatencyAverage": 380,
    "voiceLatencyAverage": 222,
    "endpointingLatencyAverage": 190,
    "numAssistantInterrupted": 0,
    "turnLatencies": [
      {
        "turnId": "t1",
        "transcriberLatency": 205,
        "endpointingLatency": 188,
        "modelFirstTokenLatency": 240,
        "modelTotalLatency": 371,
        "voiceFirstAudioLatency": 219,
        "turnLatency": 795
      }
    ]
  }
}

turnLatencies[] breaks each turn into transcriber, endpointing, model (first token and total), and voice (first audio) timings; turnLatency is the caller-perceived gap. The *Average fields roll them up. latency is null for a call that never reached the agent.

Get cost

GET /voice/calls/{call_id}/cost

What the call cost and what it consumed — what the Call Cost tab shows.

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/cost' \
  -H 'api-key: sk_your-generated-secret-here'

Response — `200 OK`

{
  "status": "success",
  "call_id": "7c2e9b4d1f6a4e3b9d0c5a8f2b1e6d47",
  "credits": { "total": 1.42, "llm": 0.91, "stt": 0.27, "tts": 0.24, "structured_output": 0.0 },
  "consumption": {
    "llm": { "input_tokens": 3120, "cached_input_tokens": 0, "output_tokens": 214, "model": "gpt-4o-mini", "provider": "openai" },
    "stt": { "duration_seconds": 51.9, "model": "nova-3", "provider": "deepgram" },
    "tts": { "character_count": 148, "model": "eleven_flash_v2_5", "provider": "elevenlabs" },
    "structured_output": { "input_tokens": 412, "output_tokens": 6, "model": "gpt-4o-mini", "provider": "openai" }
  },
  "consumption_detail": {
    "billed": {
      "llm": { "input_tokens": 3120, "output_tokens": 214, "provider": "openai", "model": "gpt-4o-mini" },
      "stt": { "duration_seconds": 51.9, "provider": "deepgram", "model": "nova-3" },
      "tts": { "character_count": 148, "provider": "elevenlabs", "model": "eleven_flash_v2_5" },
      "structured_output": { "input_tokens": 412, "output_tokens": 6, "provider": "openai", "model": "gpt-4o-mini" }
    },
    "unbilled": { "skills": {}, "customer_key": {} },
    "tts_verification": {}
  }
}

Field	Description
`credits`	The billed breakdown — `total` and its four components, the same figures as the list row's `credits` / `*_credits`. Each value is `null`, never `0`, when the billing lookup is unavailable
`consumption`	Raw usage per cost center — tokens, STT seconds, TTS characters — with the provider and model that produced it. This is the Quantity view of the Call Cost tab
`consumption_detail`	The same usage split into `billed` and `unbilled` (usage covered by your own provider keys, or by skills)

Download recording

GET /voice/calls/{call_id}/recording

The call recording as an MP3, streamed straight back to you. No link is generated and nothing is left publicly reachable — the bytes go to the caller holding the key and nowhere else.

Query parameters

Parameter	Type	Default	Description
`channel`	string	`stereo`	Which mix to return. `stereo` is the only channel produced today; `combined`, `user`, and `agent` are accepted and return `404`
`download`	boolean	`false`	`true` adds `Content-Disposition: attachment` so a browser saves the file as `recording-<call_id>-<channel>.mp3`

Headers

Header	Direction	Description
`Range`	request	Optional. Standard byte range (`bytes=0-1048575`); the response is `206 Partial Content` with `Content-Range`, so a player can seek and a download can resume
`Accept-Ranges`	response	Always `bytes`
`Content-Length`	response	The size of the body (or of the range)
`Content-Type`	response	`audio/mpeg`

curl -X GET 'https://<your-dronahq-host>/voice/calls/<call-id>/recording?channel=stereo&download=true' \
  -H 'api-key: sk_your-generated-secret-here' \
  -o recording-<call-id>.mp3

Check has_recording first

A call that never connected, or one handled by an agent with recording turned off, has no audio and returns 404 with no stereo recording for this call. has_recording on the call object tells you before you ask.

A second 404, recording not yet available, means the call did record but the upload has not landed yet — this is normal in the first minute or two after a call ends. Retry shortly.

Getting a `call_id` for a campaign call

The list above only contains dials that were actually placed — answered or not. A campaign contact whose number was invalid, who was skipped by the calling window, or whose dial could not be handed to the carrier never becomes a call, and a contact that was retried is several calls.

For campaigns, start from GET /voice/campaigns/{campaign_uuid}/attempts instead — documented with the Campaign API. It lists every dial attempt from the campaign's own registry, permanently, including the contacts that were never dialed and why, and each dialed attempt carries the call_id to fetch here.

Errors

HTTP status	Cause	Fix
`400`	Unknown `channel` on the recording endpoint	Use `stereo`
`401` / `403`	Auth failure on the API key	See the error-code chip in Request Logs (`missing_token`, `invalid_token`, `disabled`, `expired`, `scope_denied`, `agent_denied`)
`403`	The call's agent is not on the key's Agent Access list, or `agent_uuid` on the list names such an agent	Use a key that covers the agent
`404`	No such call, the call is older than 30 days, or the recording is missing	A call from another workspace is also a `404`, deliberately. For the two recording cases see Download recording
`416`	The `Range` header asks for bytes past the end of the recording	Check `Content-Length`
`422`	`start_datetime` after `end_datetime`, a bad datetime, or an unknown transcript `format`	Fix the parameter; the message names it

A call whose agent was deleted is not reachable through this API

The key's permissions are checked against the call's voice agent, so if that agent has since been deleted the request fails with 403. The call is still visible in the Call Logs screen. This fails closed on purpose — the alternative is a key reaching a call its access list can no longer be evaluated against.

Limits

Limit	Default
Calls per page	500 (default 25)
Events per page	5000 (default 1000)
Call data retention	30 days from the call's start

Phone numbers are returned as dialed

user_number, from_number, and to_number are full E.164 numbers, exactly as the Call Logs screen shows them. Treat responses from this API as personal data in your own storage and access controls.

What's Next

Campaign API — List attempts — every dial of a campaign, including contacts never called
Post-call Webhook — get pushed the same data when a call ends, instead of polling
Outbound Call Dispatch — placing the calls this API reads back
Request Logs — debugging what your integration actually sent

Call Data API

When to Use This​

Which endpoint answers which question?​

Prerequisites​

Authentication​

Endpoints​

Retention​

Pagination​

List calls​

Query parameters​

Example request​

Response — 200 OK​

The call object​

End reasons​

Get a call​

Response — 200 OK​

What it adds to the call object​

Get transcript​

Response — 200 OK​

Get events​

Query parameters​

Response — 200 OK​

Get structured output​

Response — 200 OK​

Get messages​

Response — 200 OK​

Get latency​

Response — 200 OK​

Get cost​

Response — 200 OK​

Download recording​

Query parameters​

Headers​

Getting a call_id for a campaign call​

Errors​

Limits​

What's Next​

When to Use This

Which endpoint answers which question?

Prerequisites

Authentication

Endpoints

Retention

Pagination

List calls

Query parameters

Example request

Response — `200 OK`

The call object

End reasons

Get a call

Response — `200 OK`

What it adds to the call object

Get transcript

Response — `200 OK`

Get events

Query parameters

Response — `200 OK`

Get structured output

Response — `200 OK`

Get messages

Response — `200 OK`

Get latency

Response — `200 OK`

Get cost

Response — `200 OK`

Download recording

Query parameters

Headers

Getting a `call_id` for a campaign call

Errors

Limits

What's Next