The API for AI phone calls

Send a prompt. The call happens. JSON comes back.

Start for free Read the docs

$0.05 /min all-in7 lines of code

import callingbox

call = callingbox.calls.create(

to="+15551234567",

prompt="Confirm appointment tomorrow 2pm",

context={ "patient": "Maria Lopez" },

returns={ "confirmed": "boolean" }

)

call dispatched · ai talks · json returned

+1 (680) 215-4060

Talk to a CallingBox agent right now

or listen to an example call↓

How it works

You send a request. We make the call.

Specify what you need from the call, trigger it from a POST request. You get structured JSON back.

Request

POST /v1/calls

to = "+1555..."

prompt = "Confirm appt"

returns = { confirmed }

Call

Tool call

check_availability(date="Thu")

Result

{

"confirmed":

}

Examples

What will you automate?

Write the prompt. CallingBox handles the call and returns structured data, a transcript, and a recording.

POST /v1/calls200 OK

{

"to": "+1 (555) 234-8901"

"prompt": "Confirm appointment tomorrow 2pm"

"context": { "patient": "Maria Lopez" }

"returns": { "confirmed": "bool" }

}

Same endpoint, same SDK, same webhook

0:34

Hi, calling about your 2pm appointment tomorrow with Dr. Chen.

Patient

Yes, I'll be there.

{

"confirmed": true,

"reschedule_to": null

}

POSTwebhook delivered

200 OK

Quality

The best voice quality available

Normalized quality scores across key dimensions

CallingBoxSelf-builtHuman baseline

Higher is better ↑

MOS

Turn F1

Recovery

Lower is better ↓

WER

Halluc.

Latency (ms)

Detailed results

Metric

CallingBox

Self-built

Best commercial

Human

Voice naturalness (MOS ↑)

4.31★

3.92

4.08

4.50

Response latency p50 (ms ↓)

340★

920

680

~300

Word error rate (↓)

2.9%

5.2%

4.1%

2.1%

Hallucination rate (↓)

0.9%★

8.4%

3.2%

0.5%

Turn-taking F1 (↑)

0.97★

0.81

0.89

0.99

Interruption recovery (↑)

92%★

54%

71%

96%

Evaluation on 12,400 real phone calls across US English, noisy environments (SNR 10–25dB), diverse accents. MOS rated by 3 independent evaluators per sample (ITU-T P.808). Latency measured end-to-end including network. Hallucination scored against ground-truth structured outputs. Human baseline from professional call center agents on same dataset. ★ = best non-human result. 95% CI, p < 0.001 for all CallingBox vs. self-built comparisons.

Why CallingBox

Building AI phone calls was hard and expensive. Until now.

Without CallingBox

~$0.12/min

1Server infra

SIPWebSocketsMedia

2Telephony

ProviderNumbersRouting

3Speech-to-text

StreamingChunkingVAD

4Language model

PromptsStateTools

5Text-to-speech

VoicesEncodingStreaming

6Orchestration

TurnsInterruptsErrors

7Data extraction

ParsingValidationSchema

8Operations

LogsAlertsRecovery

8 layers to build, test, and maintain

Multiple vendor contracts

Weeks of engineering

With CallingBox

$0.05/min

callingbox.create( to="+1555...", prompt="..." )

That's it.

4×

cheaper

< 500ms

latency

lines of code

Ship it today

Create an account and make your first AI phone call in minutes. No credit card required.