We replaced WebRTC with plain WebSockets for real-time GPU streaming and got lower latency, simpler infrastructure, and better reliability everywhere.

HelixML

Helix replaced WebRTC with plain WebSockets for real-time GPU streaming in their AI coding agent platform. The switch eliminated TURN server complexity, reduced latency by 20-30ms, and improved reliability across restrictive enterprise firewalls. They implemented a minimal binary protocol over WebSockets, using WebCodecs API for hardware-accelerated H264 decoding directly to canvas. Trade-offs include losing adaptive bitrate and automatic packet loss recovery, but TCP's reliability on modern networks proved sufficient. The solution works everywhere HTTPS does, requires no special firewall configuration, and provides better debuggability than WebRTC's opaque failures.

We Killed WebRTC (And Nobody Noticed)