---
title: Connect Voice AI with AgentStream
slug: exotel-agentstream/connect-voice-ai-with-agentstream
createdAt: 2026-08-11T16:03:34.497Z
---

Use Exotel [AgentStream](https://docs.exotel.com/exotel-agentstream) to stream live call audio to your bot over WebSocket, and play bot audio back to the caller.

This page gets you from zero to a working outbound test call. Pick one provider guide when you are ready to wire a full voice agent.

**API reference:** [Connect Voice AI API](https://docs.exotel.com/exotel-agentstream/connect-voice-ai-api)

## How it works

```text
Caller  ←→  Exotel  ←WSS PCM→  your bot (public wss://)
```

1. You run a bot that speaks the AgentStream events: `connected`, `start`, `media`, `stop` (and optionally `mark` / `clear`).
2. You expose that bot as a public `wss://` URL (TLS).
3. You place a call with Connect Voice AI (`StreamType=bidirectional`, `StreamUrl` = your bot).

Telephony default on AgentStream is **8 kHz, 16-bit PCM, mono**. Always pass the sample rate your bot expects on the StreamUrl query string when the bot documents it (usually `?sample-rate=8000`).

## Before you start

| You need               | Notes                                                |
| ---------------------- | ---------------------------------------------------- |
| Exotel account         | Account SID, API Key, API Token, ExoPhone (CallerId) |
| A phone number to call | E.164, for example `+91…`                            |
| Public WSS             | cloudflared, ngrok, or your own HTTPS load balancer  |
| Python 3.10+           | For the sample bots and `place_connect_call.py`      |

Clone the sample repo:

```bash
git clone https://github.com/exotel/Agent-Stream.git
cd Agent-Stream
cp shared/env.exotel.example shared/.env.exotel
# Edit shared/.env.exotel with EXOTEL_ACCOUNT_SID, EXOTEL_API_KEY,
# EXOTEL_API_TOKEN, EXOTEL_CALLER_ID
```

## Pick a provider (about 15 minutes each)

| Guide                | Mode                           | Default port | StreamUrl                                  |
| -------------------- | ------------------------------ | ------------ | ------------------------------------------ |
| [ElevenLabs](#)      | Speech-to-speech (ConvAI)      | 10002        | `wss://HOST/v1/convai/conversation/exotel` |
| [OpenAI Realtime](#) | Speech-to-speech (Realtime GA) | 5000         | `wss://HOST/?sample-rate=8000`             |
| [Sarvam](#)          | STT → TTS (Saaras / Bulbul)    | 8000         | `wss://HOST/ws?sample-rate=8000`           |

Replace `HOST` with your tunnel hostname (no `https://` prefix on StreamUrl — use `wss://`).

## Place a Connect call (shared helper)

From the repo root, after your bot and tunnel are up:

```bash
set -a && source shared/.env.exotel && set +a

python shared/place_connect_call.py \
  --to +91XXXXXXXXXX \
  --stream-url "wss://YOUR_HOST/…see table above…"
```

Endpoint used by the helper:

`POST /v1/Accounts/{AccountSid}/Calls/connect`
Auth: HTTP Basic (API Key / API Token)

| Field        | Required | Description                               |
| ------------ | -------- | ----------------------------------------- |
| `From`       | yes      | Callee in E.164                           |
| `CallerId`   | yes      | Your ExoPhone                             |
| `StreamUrl`  | yes      | Bot WSS URL (keep under \~600 characters) |
| `StreamType` | yes      | `bidirectional`                           |

## Go-live checklist (any provider)

- Bot listens on a stable host with a valid TLS certificate (not a disposable laptop tunnel for production).
- Secrets live in environment variables or a secret store — never in git.
- StreamUrl path and `sample-rate` match the provider README exactly.
- You verified pitch (not slow/deep) and at least one full user turn on a real handset.
- You read [AgentStream: WSS errors and handling](docId\:s-fw4SF8JcMzgAofPlv3b) for disconnect and timeout behaviour.

## Next

- [ElevenLabs Conversational AI](#)
- [OpenAI Realtime](#)
- [Sarvam STT/TTS](#)