You would still need some form of STUN server (but there are a number of 'freely accessible' ones, even configured as defaults in some browsers) to get your reachable address/port/proto. You cant, AFAIK, handle this manually - it is handled internally as part of ICE. Then you would need to manually handle the signalling of these as well.
Thats the bare minimum you have to do to peer over webrtc under ideal circumstances, but its certainly doable.
So you do that for each peer. If you (or a peer) change streams you will do the same thing all over again
edit: cant reply to your comment fulafel, but on that project you posted: https://github.com/cjb/serverless-webrtc/blob/master/serverl... Also, note that even if this wasnt defined, some browsers contain defaults. You 100% need STUN, but you can handle the signalling manually - as I stated.
edit2: cant reply to your comment ibc, but I was explaining the bare minimum 'serverless' webrtc case still required a STUN server. I appreciate that mediasoup SFU uses ice-lite instead.
If you say it's different with media channels, I'll believe you.
WebRTC can tell you your ip address without any external server, that only breaks if you are dealing with NAT.
edit: to clarify the NAT-less use case, i'm thinking of apps that can rely on/require p2p supporting IPv6 connectivity.