mediasoup as a Self-Hosted SFU: Cutting WebRTC Cost in Half for 2026 Voice AI

mediasoup is a C++ SFU with a Node.js control plane. For voice-only AI agents, a single 16-core box can hold thousands of concurrent peers. Self-hosting cuts the per-minute cost 4–6x versus managed clouds — if you can run it.

What it is and why now

flowchart LR
 Browser["Browser · WebRTC"] --> ICE["ICE / STUN / TURN"]
 ICE --> SFU["SFU · Pion Go gateway 1.23"]
 SFU --> NATS["NATS bus"]
 NATS --> AI["AI Worker · OpenAI Realtime"]
 AI --> NATS
 NATS --> SFU
 SFU --> Browser

mediasoup is purpose-built to route media packets — not to encode them, not to mix them, just to forward. It is faster and lighter than most general-purpose SFUs because the worker is C++; the orchestration runs in Node. In 2026 it remains the SFU of choice for teams who already operate Linux at scale and want to push managed-cloud bandwidth bills down.

The cost math is straightforward. A managed SFU minute (LiveKit Cloud, Daily, Agora) runs $0.002–0.004 per participant-minute. Self-hosted mediasoup on a $400/month bare-metal box can hold ~3,000 concurrent voice peers — a hair under $0.0005 per participant-minute including the TURN egress. Above 1,000 concurrent users the savings are real.

How WebRTC fits AI voice (architecture)

A mediasoup voice-agent setup:

1. Browser opens a WebSocket to your signaling server, fetches RTP capabilities.
2. Browser creates a SendTransport and a RecvTransport (each is a DTLS/SRTP peer connection).
3. Mic track is published as a Producer; the AI agent worker subscribes via Consumer.
4. The agent worker decodes Opus → drives STT/LLM/TTS → produces Opus back into the room.
5. Browser consumes the agent's track on the RecvTransport.

mediasoup deliberately leaves "rooms," "permissions," and "presence" to your app — you build that on top. For voice agents this is a feature: you control session lifecycle.

CallSphere implementation

CallSphere does not self-host mediasoup in production yet. We benchmarked it for a six-month window before settling on Pion + Go gateway 1.23 — the C++/Node split inside mediasoup did not pair as cleanly with our Go-everywhere services and NATS topology. That said, we recommend mediasoup to customers who want to bring their own SFU on-premises (especially behavioral-health customers under stricter data-residency rules). Across our 6 verticals, mediasoup, LiveKit OSS, and Pion all run side-by-side at customer sites without us.

For reference, our 6-container pod (CRM writer, calendar, MLS lookup, SMS, audit, transcript) sits behind whichever SFU the deployment chooses; the pod just exposes WebRTC and gRPC.

Code snippet (TypeScript, mediasoup-client)

```ts import { Device } from "mediasoup-client";

async function startCall(routerRtpCapabilities: any, signal: any) { const device = new Device(); await device.load({ routerRtpCapabilities });

const sendT = await signal.create("createWebRtcTransport"); const send = device.createSendTransport(sendT); send.on("connect", ({ dtlsParameters }, cb) => signal.send("connect", { dtlsParameters }).then(cb)); send.on("produce", (params, cb) => signal.send("produce", params).then(({ id }: any) => cb({ id })));

const stream = await navigator.mediaDevices.getUserMedia({ audio: true }); await send.produce({ track: stream.getAudioTracks()[0] });

const recvT = await signal.create("createWebRtcTransport"); const recv = device.createRecvTransport(recvT); // subscribe to agent producer via signaling, then recv.consume(...) } ```

Build / migration steps

1. Spin up a Linux box with public IPs for ICE; install Node 20 + Python 3 (for mediasoup workers).
2. Implement signaling (Socket.IO works well) for transport, producer, and consumer events.
3. Run one mediasoup worker per CPU core; bind `rtcMinPort`/`rtcMaxPort` to a 10k UDP range.
4. Deploy a co-located AI worker that consumes the user producer, streams Opus to your STT, and produces synthesized Opus back.
5. Stand up coturn for fallback relay; expose 443/TLS for restrictive networks.
6. Monitor with the mediasoup observer events; alert on transport `iceState === "disconnected"` > 5 s.

FAQ

Is mediasoup OSS friendly to commercial use? Yes — ISC license. How many cores per 1k voice peers? Roughly 4 cores at 16 kbps Opus once tuned. Does it support SVC/simulcast? Yes for video; for voice, you stay in single-stream Opus. Can it work with OpenAI Realtime? Yes — bridge in your AI worker process; the worker maintains a Realtime session per call. What about end-to-end encryption? SRTP terminates at the SFU; for E2EE add Insertable Streams.

Sources

• https://mediasoup.org/
• https://github.com/versatica/mediasoup
• https://webrtc.ventures/2022/05/webrtc-with-mediasoup/
• https://openvidu.io/3.1.0/docs/self-hosting/production-ready/performance/

Explore a live demo and compare current plans to find the right fit for your business.

mediasoup as a Self-Hosted SFU: Cutting WebRTC Cost in Half for 2026 Voice AI: production view

mediasoup as a Self-Hosted SFU: Cutting WebRTC Cost in Half for 2026 Voice AI sits on top of a regional VPC and a cold-start problem you only see at 3am. If your voice stack lives in us-east-1 but your customer is calling from a Sydney mobile network, the round-trip time alone wrecks turn-taking. Multi-region routing, GPU residency, and warm pools become the difference between "natural" and "robotic" — and it's all infra, not the model.

Serving stack tradeoffs

The big fork is managed (OpenAI Realtime, ElevenLabs Conversational AI) versus self-hosted on GPUs you operate. Managed wins on cold-start, model freshness, and zero-ops; self-hosted wins on unit economics past a certain conversation volume and on data residency for regulated verticals. CallSphere runs hybrid: Realtime for live calls, self-hosted Whisper + a hosted LLM for async, both routed through a Go gateway that enforces per-tenant rate limits.

Latency budgets are non-negotiable on voice. End-to-end target is sub-800ms ASR-to-first-token and sub-1.4s first-audio-out; anything beyond that and turn-taking feels stilted. GPU residency in the same region as your TURN servers matters more than choosing a slightly bigger model.

Observability is the unglamorous backbone — every conversation produces logs, traces, sentiment scoring, and cost attribution piped to a per-tenant dashboard. HIPAA aligned isolation keeps healthcare traffic separated from salon traffic at the storage layer, not just the API.

FAQ

Is this realistic for a small business, or is it enterprise-only? The IT Helpdesk product is built on ChromaDB for RAG over runbooks, Supabase for auth and storage, and 40+ data models covering tickets, assets, MSP clients, and escalation chains. For a topic like "mediasoup as a Self-Hosted SFU: Cutting WebRTC Cost in Half for 2026 Voice AI", that means you're not starting from scratch — you're configuring an agent template that's already been hardened across thousands of conversations.

Which integrations have to be in place before launch? Day one is integration mapping (scheduler, CRM, messaging) and prompt tuning against your top 20 real call transcripts. Day two through five is shadow-mode running, where the agent transcribes and recommends but a human still answers, so you can compare side-by-side. Go-live is the moment your eval pass-rate clears your internal bar.

How do we measure whether it's actually working? The honest answer: it scales until your tool catalog gets stale. The agent is only as good as the integrations it can actually call, so the operational discipline is keeping schemas, webhooks, and fallback paths green. The platform handles the rest — observability, retries, multi-region routing — without your team owning the GPU layer.

Talk to us

Explore a live demo and compare current plans to find the right fit for your business.