Closio for Shopify
Voice mode
Voice mode lets a shopper talk to the agent and hear the answer, with live captions. It uses the same tools and the same guardrails as chat: the agent proposes items aloud and adds them only after a clear yes. Voice is included on Growth (200 minutes a month) and Pro (1,000 minutes).
Updated 2026-11-06
How a call works
The shopper's browser connects to the speech model over WebRTC. The connection is set up by Closio's server, which holds the API key; the browser never sees a credential. Every action the model wants to take (look up stock, propose a cart, check an offer) comes back through Shopify's signed app proxy and runs with the same code as chat.
If the live service is unavailable, or the microphone or WebRTC fail, the widget falls back to the browser's own speech recognition and synthesis through the chat endpoint. The shopper sees the same orb and captions.
Default model gpt-realtime-2.1-mini with gpt-4o-mini-transcribe for captions; both can be changed by Closio without a release.
Minutes, credits and what happens at zero
Minutes are counted on the server from wall-clock time (never less than what the widget reports), per 30-day usage period. The plan allowance is used first, then purchased voice credits. A call is never cut mid-sentence: each call gets a cap when it starts (at most 30 minutes, or your Controls limit), and the cap shrinks to what is left in the budget.
When plan minutes and credits are both at zero, a new call is refused with "Voice minutes are used up for now. You can keep chatting by text." and the shopper stays in the text chat. Nothing changes for chats already open.
- Abandoned tabs stop billing: the widget sends a heartbeat every 45 seconds; a call without one for 3 minutes is closed and billed only for the time it was alive.
- At most 3 simultaneous calls per store; a fourth hears that the lines are busy.
Controls you have
Controls → Voice:
- Voice name (10 voices; default "marin").
- Barge-in: whether the shopper's speech interrupts the agent (default on).
- Store transcripts: off means no voice transcript is saved, only the minutes.
- Longest call (1–60 minutes), calls per day per store, calls per day per shopper.
- Working hours for voice, separate from chat.
Privacy
Audio is processed live and is not stored by Closio. When transcript storage is on, the transcript is saved as messages of the conversation after redaction: emails, phone numbers, card and IBAN numbers and (for the languages below) street addresses are masked before storage. A one-line call summary is saved for the Inbox.
Language coverage
The agent replies in the shopper's language in chat; Controls can restrict replies to a list and set a fallback language. For voice, Closio has verified English on a live store. Other languages are supported by the speech model according to its provider, but Closio has not yet tested them end to end; treat them as available but unverified and test with your own store before promoting voice in that language.
| Language | Chat (text) | Live voice (OpenAI Realtime) | Browser voice fallback | Transcript redaction |
|---|---|---|---|---|
| English | Yes, verified | Yes, verified on a live store | Depends on browser/OS | Email, phone, card, IBAN, street address |
| French | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN, street address |
| German | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN, street address |
| Spanish | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN, street address |
| Italian | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Portuguese | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Dutch | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Japanese | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Chinese | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Arabic (RTL layout) | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Hebrew (RTL layout) | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
| Swedish, Danish, Norwegian, Finnish, Polish, Czech, Turkish, Korean, Russian, Ukrainian, Greek, Romanian, Hungarian, Thai, Vietnamese, Indonesian, Hindi | Yes | Supported by the model; not yet verified by us | Depends on browser/OS | Email, phone, card, IBAN |
The widget mirrors its layout for right-to-left languages (Arabic, Hebrew, Persian, Urdu). Browser fallback quality varies by browser and operating system; Chrome on desktop has the broadest coverage.
Latency and health
Settings → Status shows connect attempts, connect success, connect time (median and slowest 5 %) and time to first audio, for the last 24 hours and 7 days, with a plain verdict. Voice failures can also notify you by Slack, webhook or email.