Voice agents

An AI voice agent gateway for legacy SIP

Your switch speaks SIP over UDP with plain RTP. OpenAI's GPT-Live-1 wants TLS, SRTP, a webhook and a live session. Point the switch at Open Voice Shield and the two talk.

Open Voice Shield sits between a SIP switch, SBC or PBX and an AI voice agent. The switch sends an ordinary call over UDP or TCP. The platform handles TLS, encrypted media, the webhook and the agent's session, and hands the switch a plain one-hop dialog. The call is recorded, transcribed and scored like any other, and the agent leg is metered per second.

The gap

What an agent expects, and what a switch can do

Voice agents are built for the web. OpenAI exposes GPT-Live-1 over SIP, but the contract is a web contract: TLS signalling on port 5061, SRTP media, an HTTPS webhook that must accept each incoming call with the model, voice and instructions, and a session the agent only runs while a client is attached to it. xAI's agents want TLS and a digest credential. None of that exists in a class-4 softswitch, most SBCs, or a PBX that has been in service for a decade.

Where a legacy switch stops and the platform takes over.
The agent expectsYour switch hasThe platform does
TLS signallingUDP or TCP on 5060Terminates plain SIP, originates TLS to the agent
SRTP mediaPlain RTPEncrypts the agent leg; your leg stays plain
Telephony codecs onlyG.711, sometimes G.729Offers the agent G.711 and G.722, nothing else
A webhook that accepts the callNothingHosts one per destination, signature-checked
A session held for the callNothingAttaches and holds the agent's control connection
A dialog it can route back toOften ignores route setsOne plain hop and a Contact that is the platform

How it works

One route on the switch, one destination on the platform

  1. Your switch sends the call to the ingress, over UDP or TCP with an early SDP offer, authorized by its IP address, as for any other route.
  2. A routing rule picks a Voice AI destination: the agent provider, the project or agent identifier, and the agent's voice and instructions.
  3. The platform originates the agent leg: TLS signalling, SRTP media, telephony codecs. For OpenAI, it accepts the call through the webhook it hosts and attaches the session. For xAI, it answers the digest challenge.
  4. The switch gets a normal answer: plain RTP, a single route hop, a Contact on the platform. Its ACK, re-INVITEs and BYE are re-targeted to the agent.
  5. The call is relayed, recorded and analyzed, and when either side hangs up, the other side is told at once.

The switch never sees a certificate, a crypto attribute, a webhook or a session. From its side, the agent is a carrier route that happens to answer every call.

Who it is for

Carriers, contact centres and PBX operators

Carriers and ITSPs
Offer AI answering, screening or after-hours agents to wholesale and retail customers from the softswitch you already run. Each customer gets their own destination, secrets and per-second metering, and you set the retail rate.
Contact centres
Put a GPT-Live-1 agent in front of a queue, a DID or an overflow route without touching the PBX's transport stack. Every agent conversation comes back recorded, transcribed and scored.
PBX and SBC operators
Divert flagged or unwanted callers to an agent instead of a human, or route a prefix to an agent for a pilot. The same platform screens the traffic for fraud.

Supported agents

Which agents, and what you get on every call

Today: OpenAI's GPT-Live-1, the first voice model that behaves like a live conversation partner, the OpenAI Realtime models, and xAI's voice agents. Each is a Voice AI destination with its own connection rules. The list is published by the API and in the portal, and grows as providers are tested end to end.

  • Stereo recording, caller on one channel, agent on the other.
  • Transcript, verdict and red flags, as for a human-answered call.
  • The SIP ladder and per-leg quality, with estimated MOS.
  • The agent leg metered per second, shown on the record and on the invoice.
  • The agent's own transcript of what it heard and said, on the destination.

Each destination has a call handling setting: full analysis, record only, or relay only. Relay only skips the recording, transcript, verdict and SIP log. The call is bridged and metered, and the record keeps the facts of the call. The default stays full analysis.

When the agent asks to transfer the caller, the platform checks the number against your account and places the new call itself through your routing, connects the caller to it and releases the agent. Your switch sees only the same call. The new leg is routed, recorded and billed like any other, and the call record shows the transfer, how far it got and the new call. A switch that handles transfers itself can be handed the SIP REFER instead, per destination. You can start a transfer yourself from the call page, the API or an MCP tool, and the keys a caller presses are on the record. An agent-initiated hang-up ends the call for the switch at once.

Get started

Connect a switch to an agent

The connect any SBC to a voice agent guide takes a switch from nothing to a live agent call in four steps: whitelist the switch, add the destination, register the webhook, route a prefix. Provider setup details are in the SIP integration guide. Or start free with $5 credit and dial an agent from the web dialer first.

Questions

Frequently asked questions

Can a SIP switch that only does UDP and plain RTP talk to an AI voice agent?

Yes, through Open Voice Shield. The switch sends an ordinary SIP call over UDP or TCP with plain RTP to the ingress. The platform connects to the agent provider over TLS with SRTP media, hosts the webhook that accepts the call, holds the agent's session for the whole call, and presents the switch a normal one-hop call it can answer, acknowledge and hang up. Nothing changes on the switch beyond a route.

Which AI voice agents are supported?

OpenAI's GPT-Live-1 and the Realtime models, and xAI's voice agents. Each is a Voice AI destination with its own connection rules. The provider list is published by the API and in the portal.

Why can't my SBC connect to OpenAI directly?

OpenAI's SIP endpoint requires TLS signalling and SRTP media, a call is only connected when a webhook you host accepts it with a model, voice and instructions, and a GPT-Live agent only speaks while a control session is attached. Most switches and SBCs cannot do any of those. The platform does all of them per destination.

Does the switch need any change for encryption?

No. The switch keeps plain SIP and plain RTP. The platform encrypts the leg towards the agent, and the answer it sends the switch carries one plain route hop and a Contact that is the platform, so even a switch that ignores route sets or refuses secure URIs completes the call.

Is a call to an agent recorded and analyzed like other calls?

Yes. Both legs are relayed and recorded in stereo, transcribed and scored, with the SIP ladder and per-leg quality on the call record. The agent leg is metered per second at a rate the operator sets.

What happens when the agent hangs up or transfers?

When the agent ends the session, the platform ends the call towards the switch at once. When the agent transfers the caller, the platform places the new call itself through your routing, connects the caller to it and releases the agent, so a switch that cannot take a SIP REFER still gets the transfer. You can also start a transfer yourself from the call page, the API or an MCP tool. Keys the caller presses are on the call record.