Most “voice AI” tutorials hand you a managed platform, a magic SDK, and a bill that scales with per-participant minutes. This post is about the other path: wiring a Pipecat bot to the raw Cloudflare Realtime SFU over its new WebSocket media adapter, paying only for bandwidth, and owning the whole media path.
I'll walk through the architecture I built, what I learned along the way, the uncomfortable fact that the adapter is beta, the production architecture I'd actually ship, and then a grounded comparison against LiveKit — the obvious alternative — on the two axes that decide these projects: latency and cost. Because LiveKit is open source, I treat it as two options throughout: LiveKit Cloud (managed) and self-hosted LiveKit (you run the SFU).
