We picked WebRTC for peer-to-peer port sharing at 3:05 in the morning. By 9:41 we had switched to a relay

The plan was no server of ours. WebRTC between the two Macs, introductions over public Nostr relays, public STUN. We decided on it at 3:05.

At 3:24 the research came back. Four desk reviews, every claim labeled documented, reported or inferred:

• Direct connections with STUN only would work for an estimated 75 to 88% of people (inferred). The rest are on networks that block UDP, like offices and guest Wi-Fi, which is where people open a shared link.
• Public Nostr relays are turning this traffic away. strfry, which runs the largest ones, ships a plugin to reject relayed WebRTC signaling (documented).
• We found no browser P2P product that runs without its own relay and publishes its success rate.

So at 9:41 we went relay first. Both Macs connect outward to a small relay over a secure WebSocket on port 443, which works on every network. The session is Noise IK end to end, so the relay forwards bytes it can't read. It stores nothing: one small program with no database, which we plan to package so anyone can run their own. A direct path can come later behind the same seam, the way Tailscale starts on its relays.

It was carrying traffic between two of our machines by 17:36.

03:05  WebRTC is the transport
03:24  research: gotchas in WebRTC, Nostr and port mapping
09:41  relay-first architecture spec
13:25  the relay and Noise IK, deployed
17:36  live: two instances through relay1.port42.ai