How Speakvora agents work
From a description of your business to a published agent with an endpoint: what happens in a turn, what the agent can do, and exactly where it stops.
An agent is a published version
You describe the business in the builder (or paste a website address and review the draft it writes). That produces a configuration: name, purpose, instructions, personality, language, voice, business knowledge, tools, appointment settings and which channels are on. Publishing saves it as a numbered, unchangeable version and gives you an endpoint and a key that works for that one agent. Change something later and you publish version 2; version 1 keeps working until you replace it.
What happens in one turn
- Input. Your app sends text, or audio that is transcribed first.
- Reply. The language model answers using only the business knowledge you gave it for prices, hours and policies; if it does not know, it says so and offers to take a message.
- Tools. If the person asks for something an agent can do (check times, book, move or cancel, send a deposit link, take a message, check an order status through your own endpoint), it calls the tool and uses the result. Only the results of a tool are used for facts; it never invents times or prices.
- Speech. The reply is split into sentences. Each sentence is spoken as soon as it is ready: our phrase library first (instant), then live generation.
- Memory. The conversation is kept for the session so the next turn has context. If the listener interrupts, memory is cut back to what they actually heard.
The response gives you the text and an audio link for every sentence, plus timings:
curl -X POST https://speakvora.com/api/v1/agents/AGENT_ID/turn \
-H "x-api-key: YOUR_AGENT_KEY" -H "content-type: application/json" \
-d '{"text":"Do you have anything open Tuesday morning?","session":"visitor-123"}'
# 200 OK
{"sentences":[{"index":0,"text":"Yes, we have 9:00, 9:30 and 10:00 on Tuesday morning.","audio_url":"https://speakvora.com/live/.../....mp3","source":"live"}],
"reply":"Yes, we have 9:00, 9:30 and 10:00 on Tuesday morning.","session":"visitor-123","timings":{"first_text_ms":520,"llm_ms":480,"total_ms":900}}Channels: same agent, same facts
- Your app or website: call the turn endpoint, or paste one script tag for a text chat widget that makes no audio.
- Phone and texts: connect your own Twilio or Telnyx account in the dashboard (your provider bills you directly); calls are turn-based with a short pause, not a live stream.
- Email: the agent drafts replies from the same knowledge for you to approve; plain questions can be answered automatically if you switch that on.
- Reminders and a daily summary: appointment reminders, no-show follow-ups, review requests and a daily email of bookings and unanswered questions.
Languages and voices
Pick a language and a catalog voice, use your own recorded voice, or set the language to automatic and the agent answers in the language the customer used (English, Spanish, French, Portuguese, Italian, Hindi, Japanese or Chinese).
What it costs
$10 per 1,000 tasks, with 200 free. A task is one handled message of any kind: a voice turn, a chat message, a text, an email reply or an automated reminder. Booking, deposits, email, chat and reminders are included in that one price.
Honest limits
- It gathers and passes on information. It cannot refund, charge or change your systems, except the appointment actions its tools provide.
- Phone calls are turn-based, not full-duplex. There is a pause after the person stops talking.
- It says it is an AI assistant when asked and gives no medical advice; if someone describes an emergency it tells them to call emergency services.
- Not HIPAA-ready: do not use it for patient information.