The container's own startup line is not an address to dial: it names the interface
and port the process bound inside the container, ws://0.0.0.0:8080/session, and
0.0.0.0 is not an address to dial. Nor is it the ws://127.0.0.1:8080 that
vscode_client's first session names for a server on this machine: 8080 on this
machine is the card's page container while the pair is up, and that page relays the
session endpoint to the card's server, so an editor pointed at 8080 hosts on the
card's server rather than this one — and answers nothing when the pair is down.
This container serves the session protocol and no page, so a client is what
joins it; for a room you can open in a browser, run the two containers the Run
card prints, at https://selvage.dontblameme.dev/#try. The page image is
web_client's and its README owns that image's configuration and its tags.
The default binds loopback, so only your own machine reaches it. For a friend to join, bind an address they can reach and hand them a URL that names your machine:
nix develop . -c cargo run -p selvaged -- --listen 0.0.0.0:8080A free tunnel that forwards to your port works for a first test. Beyond that you want a machine with a public address (a small VPS, and the right firewall rules).
Plain ws:// and http:// is plaintext: the invite token travels in the clear,
so it is for a tailnet, a VPN or loopback. The protocol's transport security is
the deployer's to supply; put a TLS terminator in front (tailscale serve,
caddy, or your edge) and hand out the https:// or wss:// URL. There is no
TLS inside selvaged, and none is claimed.
The image is built multi-arch, so an arm64 host pulls the same tag as an amd64
one, and it carries the server alone.
compose.yaml at the root is the self-host shape, and
docker compose up runs the server beside the published page image, both
read-only with every capability dropped.