Everything underneath a voice agent, built as one system.
Xebit is nine components that ship as one platform: Agent SDK, Bridge, Agent Server, speech pipeline, orchestration, scheduling and campaigns, observability, control plane and SIP gateway. The same artefact runs on Docker or Kubernetes in our cloud and in yours. Point vendors sell you one layer and leave the integration to you, which is where the latency, the cost and the three in the morning debugging live.
What are the components of the Xebit stack?
- 01
Agent SDK
What your team builds agents against.
- 02
Bridge · SFU
Real time audio routing and fan out across SIP, WebRTC, WebSocket and gRPC, so a person can join an AI call from a browser. Sets your cost per node.
- 03
Agent Server
Agent and model session logic, on its own resource profile.
- 04
Speech pipeline
Speech to speech by default. Cascade when you want to pick each component.
- 05
Orchestration
Turn handling, barge in, business rules enforced in the call flow.
- 06
Scheduling & campaigns
Dialling plans, retries, list management, pacing.
- 07
Observability
Per call and per node metrics, transcripts, event logs, replay, and an SRE agent that opens incidents, writes the RCA and proposes a fix for you to approve.
- 08
Control plane
Auth, tokens, dispatch. In front of the media path, not inside it.
- 09
SIP gateway
Your carrier trunks, over standard SIP and WebSocket.
What path does a call take?
- 01
Your carrier
PSTN call arrives on trunks you already own
- 02
SIP gateway
Standard SIP and WebSocket ingress
- 03
Bridge · SFU
Real time audio routing for the session
- 04
Agent Server
Agent and model session logic
- 05
Speech pipeline
Speech to speech by default, or cascade
Which protocols does Xebit support?
Mix protocols on one call: a customer rings in over SIP, the AI agent answers, and a team lead listens in from a browser over WebRTC. One session, one Bridge, audio fanned out to everyone in real time.
In production
RTP over SIP
Phone callers
The telephone network's own language. Your carrier trunks and PBX speak it today, so calls arrive without a middleman.
Best for
- Inbound and outbound PSTN calls
- Your existing carrier and numbers
- Contact centre and PBX transfers
SIP signalling · RTP media · G.711 / Opus
WebRTC
Browser and app users
Real time audio straight from a browser or mobile app, with echo cancellation and jitter handling built in.
Best for
- Talk buttons on your website and app
- A person joining an AI call from a browser
- Supervisors listening in live
UDP · SRTP · Opus
WebSocket
Platforms and servers
Audio frames over one persistent connection. The simplest way for a server, or a telephony platform that streams call audio, to plug in.
Best for
- Telephony platforms that stream call audio
- Server to server audio
- Custom clients and quick prototypes
TCP · TLS · PCM or µ-law frames
gRPC
Your backend services
Typed, bidirectional streaming between services. Built for internal pipelines that need a strict contract and high throughput.
Best for
- Backend to backend audio pipelines
- Analytics and transcription taps
- Service meshes inside your network
HTTP/2 · bidirectional streams · Protobuf
On the roadmap
AudioSocket
A lean TCP audio protocol from Asterisk: connect an existing Asterisk PBX without a SIP trunk.
QUIC
UDP transport with TLS 1.3 built in and no head-of-line blocking; holds up when mobile networks switch.
MoQ
Media over QUIC, the emerging IETF standard for low latency live media fan-out at scale.
Where does audio go in each deployment model?
Where audio goes: Audio leaves your network: it reaches Xebit's infrastructure and the model provider.
Where audio goes: The platform stays inside your network. Inference traffic goes out to your model provider.
Where audio goes: No route out of the network. Models run locally on GPUs you provide.