An AI agent and call insights, on the models you choose
A natural-sounding voice agent that answers with your greeting and rules, and transcripts, summaries, topics and mood for every recorded call. Use a provider's API, or keep it all on your own servers.
What the AI agent does
It joins the call like any other destination in your phone system, and speaks for your organization.
Place it anywhere in the call flow
The AI agent is a block in the IVR builder and a destination for any phone number. Send callers to it when the office is closed, offer it as a menu option, or give it a number of its own. Queues and voicemail stay one key away.
- After hours, from the business-hours block
- As a key in any menu
- Directly on a phone number
- Every AI call shows on the Live page and in analytics
Each organization picks its own providers, keys and voice
Choose the speech provider and the language model from a dropdown, paste your keys, pick a voice, and write the greeting and instructions. Test the connection from the same page. Another organization on the same installation can make completely different choices.
| What | Choose from |
|---|---|
| Speech: ears and voice | Deepgram, or your own speech server |
| Language model | OpenAI, Google Gemini, Anthropic Claude, or any OpenAI-compatible server |
Your own speech and language models, through two simple endpoints
Banks and hospitals often can't send call audio to an outside cloud. Point the platform at your own server instead: it only has to answer two HTTPS requests.
POST {your-url}/stt
Authorization: Bearer <token>
Content-Type: audio/wav // PCM16, mono, 8 kHz
→ 200 { "text": "My card got blocked after I travelled.",
"segments": [{ "start": 0.0, "end": 2.1, "text": "…" }] }POST {your-url}/tts
Authorization: Bearer <token>
{ "text": "I can unblock it after a quick check.",
"voice": "your-voice-name", "sample_rate": 8000 }
→ 200 audio/wav // PCM16, mono, 8 kHzFor the language model, any server that speaks the OpenAI chat format works, with or without a key. The connector guide has a complete example server and the exact request and response formats.
Read the connector guideEvery recorded call, summarised and searchable
After a call ends, a job is queued to transcribe and analyse it. Jobs run in the background across all your servers, are retried if a provider is down, and are listed on the Jobs page. Results appear on the call and on the Insights page, and fire webhooks.
- transcript.ready and call.analyzed webhooks
- keyword.detected when a watched word is said
- Turn analysis on or off per organization
You decide where call data goes
Questions about the AI
The connector guide and security overview go deeper.
Let the AI agent take the night shift
Create an organization, add your AI keys or your own server, and put the agent on your after-hours path.