# OpenAI Realtime

Run an OpenAI Realtime voice agent on Vonyx numbers through a LiveKit bridge.

OpenAI's SIP endpoint only accepts TLS signaling with encrypted (SRTP) audio. Vonyx trunks use UDP or TCP without media encryption, so you can't point a Vonyx trunk straight at OpenAI today. Instead, run the OpenAI Realtime model inside a LiveKit agent, and connect LiveKit to Vonyx over SIP.

<Callout type="warn" title="A direct SIP connection doesn't work">
OpenAI's SIP URI is `sip:$PROJECT_ID@sip.api.openai.com;transport=tls`. OpenAI requires TLS signaling on port 5061 and SRTP for call audio. Vonyx supports neither, so calls sent from a Vonyx trunk to that URI will fail.
</Callout>

## How it connects

| Leg | Between | Protocol |
| --- | --- | --- |
| Phone call | Caller and Vonyx | PSTN |
| SIP trunk | Vonyx and LiveKit | SIP over UDP or TCP, PCMU / PCMA |
| Model | LiveKit agent and OpenAI | Realtime API, using your `OPENAI_API_KEY` |

The Realtime model still handles speech-to-speech. LiveKit carries the call audio between Vonyx and the model.

## Before you start

- A Vonyx [SIP trunk](/docs/console/sip-trunks) with a [number](/docs/console/numbers) attached
- A LiveKit Cloud project and the LiveKit CLI
- An OpenAI API key

## Set up the bridge

<Steps>
<Step>

### Connect Vonyx to LiveKit

Follow the [LiveKit guide](/docs/integrations/livekit) to create the outbound trunk, the inbound trunk and the dispatch rule, and to set the Vonyx inbound address. That guide covers both outbound and inbound calls.

</Step>
<Step>

### Use the OpenAI Realtime model in your agent

Install the LiveKit Agents OpenAI plugin:

```bash
uv add "livekit-agents[openai]~=1.8"
```

Set `OPENAI_API_KEY` in your agent's `.env`, then pass the Realtime model as the session's LLM:

```python title="agent.py"
from livekit.agents import AgentSession
from livekit.plugins import openai

session = AgentSession(
    llm=openai.realtime.RealtimeModel(voice="marin"),
)
```

OpenAI bills the model usage to your key.

</Step>
</Steps>

<Callout title="Using Pipecat instead">
Pipecat ships an OpenAI Realtime service (`OpenAIRealtimeLLMService`, installed with `uv add "pipecat-ai[openai]"`). You still need a SIP transport in front of Pipecat that accepts UDP or TCP from Vonyx.
</Callout>

## Test the connection

1. Start your LiveKit agent.
2. Place an outbound call as shown in the [LiveKit guide](/docs/integrations/livekit#outbound-calls). Your phone should ring and the Realtime agent should answer.
3. Call your Vonyx number from a phone. The call should reach a `call-` room and the agent should pick up.

## Limitations

- DTMF and call transfer are not available on Vonyx yet. OpenAI's `refer` endpoint for transfers doesn't apply on this setup.
- New accounts make 1 call at a time ([raise limits](/docs/sip/limits)).
- Vonyx can't connect to OpenAI's SIP endpoint directly, because it has no TLS or SRTP support.
- The Realtime API's call webhooks (`realtime.call.incoming`) and the `/v1/realtime/calls/{call_id}/accept`, `/reject` and `/hangup` endpoints are only for OpenAI SIP calls. They aren't used when the call goes through LiveKit.

## Troubleshooting

| Problem | Fix |
| --- | --- |
| Calls to `sip.api.openai.com` fail or time out | Expected: OpenAI needs TLS and SRTP. Use the LiveKit bridge above. |
| The call connects but the agent is silent | Check that `OPENAI_API_KEY` is set where the agent runs and that the agent joined the room. |
| Inbound calls don't reach LiveKit | Check the inbound address in the Vonyx console and the inbound trunk's allowed address. See [troubleshooting](/docs/sip/troubleshooting). |
| Outbound calls are rejected | Make sure the caller ID is one of your Vonyx numbers and the trunk credentials match. |