Two control surfaces over one censorship-resistant transport:
| Leg | What | Port on PC |
|---|---|---|
| Browser | omp collab guest SPA (transcript, prompt, interrupt, subagents) | 7466 |
| Browser | Harness Remote PWA (separate pre-existing control plane) | 5173 |
| Terminal | user-mode sshd -> tmux attach -t omp |
2222 |
Phone -> PC transport: iroh-ssh (QUIC/UDP with HTTPS-443 relay fallback, E2E encrypted).
The PC's endpoint id is derived from ~/.ssh/irohssh_ed25519 and is stable across reboots.
~/Code/omp-remote is a git repo holding every file needed to rebuild the runtime dir,
including the vendored SPA bundles and both control-plane legs (omp collab + Harness Remote).
The runtime dir ~/.local/share/omp-remote/ is
disposable and has been deleted once already (2026-09-29 18:18-18:21, by a parallel agent's
"orphan cleanup" that also killed the tmux server), which broke the deployment silently.
~/Code/omp-remote/install.sh # rebuild/repair the runtime dir and the units| Path | Role |
|---|---|
~/.local/share/omp-remote/collab-serve.ts |
self-hosted collab relay + SPA, single origin (127.0.0.1:7466) |
~/.local/share/omp-remote/dist/ |
collab-web SPA (production bundle) |
~/.local/share/omp-remote/harness-web/ |
Harness Remote PWA bundle + its static server (serve.ts) |
~/.local/share/omp-remote/collab-overlay.yml |
omp overlay: collab.autoStart=control, relayUrl=ws://127.0.0.1:7466 |
~/.local/share/omp-remote/sshd/sshd_config |
root-free sshd, loopback 2222, key-only |
~/.local/share/omp-remote/iroh-serve.sh |
iroh-ssh server launcher (needs an unshare -rm mount namespace) |
~/.local/share/omp-remote/phone-link.sh |
prints the guest link rewritten for the phone's forward port |
omp-sshd.service— user-mode sshd on 127.0.0.1:2222omp-iroh-ssh.service— iroh-ssh server (endpoint id below)omp-collab-relay.service— relay + SPA on 127.0.0.1:7466harness-remote.service— Harness Remote gateway on 127.0.0.1:4900 (Basic Auth; lists and resumes native omp sessions across Projects — a different surface from the collab room, which mirrors one live session). Needsharness-remoteon PATH (npm i -g harness-remote).harness-remote-web.service— Harness Remote PWA on 127 8CA3 .0.0.1:5173 (phone-forwarded as 8444)omp-tmux.service— ensures theomptmux session exists at boot. It starts a dedicated session (no--continue): resuming "the most recently written session" was observed attaching to a session that two other live omp processes were already writing. Continue an older conversation from inside the session (/resume) or over the SSH leg.KillMode=processkeeps a unit restart from killing the tmux server.
unshare -rm is required for iroh-ssh because this host blackholes IPv6 loopback: /etc/hosts
and /etc/gai.conf are bind-mounted inside the namespace with ::1 removed and IPv4 precedence
added, so is_ssh_server_available("localhost:2222") can resolve. No sudo is needed.
tmux new -As omp # persistent session (linger is on)
omp --config ~/.local/share/omp-remote/collab-overlay.yml
# then, from anywhere:
omp collab list --json # instanceId / participants / relayConnected
~/.local/share/omp-remote/phone-link.sh 8443 # link to open on the phonepkg update && pkg install -y rust git openssh
./omp-phone.sh prepare # builds the patched iroh-ssh (10-20 min on a phone)
./omp-phone.sh key # prints the PUBLIC key to authorize on the PC
cat ~/.ssh/id_ed25519_omp.pub # append this to the PC's ~/.ssh/authorized_keysDaily use — one process gives every surface (the endpoint id is not in this repo; see below):
mkdir -p ~/.config/omp-remote
echo '<your-endpoint-id>' > ~/.config/omp-remote/endpoint # once, from 'iroh-ssh info' on the PC
./omp-phone.sh test # end-to-end check
./omp-phone.sh connect # persistent tunnel, auto-reconnects
# browser: http://127.0.0.1:8443/ <- link printed by phone-link.sh
# PWA: http://127.0.0.1:8444/ (add machine 127.0.0.1:4900, creds from harness-remote.env)
# terminal: ssh -i ~/.ssh/id_ed25519_omp -p 2222 kir@127.0.0.1 -t 'tmux attach -t omp'Equivalent by hand:
iroh-ssh -N \
-o IdentityFile=~/.ssh/id_ed25519_omp -o IdentitiesOnly=yes \
-L 8443:127.0.0.1:7466 -L 8444:127.0.0.1:5173 -L 4900:127.0.0.1:4900 -L 2222:127.0.0.1:2222 \
kir@<your-endpoint-id> &- The guest link's room id/key changes per session; only the
127.0.0.1:7466part is rewritten. 127.0.0.1is a secure context, so the SPA gets WebCrypto over plain HTTP — no certificates.- The iroh relay fallback rides n0's shared public relays (stateless, E2E encrypted) only when a direct QUIC path cannot be established; nothing is hosted by a third party besides that relay.
- Key-only auth: the endpoint id is the only secret needed to reach port 2222, so keep it private.
- The PWA needs the gateway's Basic Auth once per device: add a machine with host
127.0.0.1, port4900(the forward target, not 8444) and the credentials from~/.config/omp-remote/harness-remote.env(usernameharness). The browser then remembers it. - The gateway only accepts browser calls from origins listed with
--cors. The tunnel maps the PWA to a different port than the API, which makes every call cross-origin, so bothhttp://127.0.0.1:8444andhttp://localhost:8444are allow-listed in the unit. Open the PWA on one of those two origins (not the LAN IP) or the machine shows as unreachable. - Termux has no
netstat;omp-phone.sh diagreads/proc/net/tcpinstead.
The phone builds a patched iroh-ssh (cargo install --git https://github.com/Kiraprint/iroh-ssh).
The change that matters on mobile:
Dial the peer through its home relay URL — EndpointAddr::new(id).with_relay_url(..) instead of
a bare endpoint id — so a connection never depends on iroh discovery.
Why: stock iroh-ssh can only find a peer through iroh's discovery, which resolves the record
_iroh.<z32-endpoint-id>.dns.iroh.link over the system resolver. On Android there is no
/etc/resolv.conf for Termux to read, so iroh falls back to hardcoded public resolvers over plain
UDP/53 — the first thing carrier networks block, intercept or answer from a stale cache. The client
then dies with:
Error: No addressing information available
1: Discovery produced no results for <endpoint id>
Dialling the peer's relay directly removes discovery from the path entirely.
Correction, worth knowing if you read the earlier version of this note: an earlier analysis here claimed n0's DNS zone was stale, based on
NXDOMAINfor_iroh4.<z32>.dns.iroh.link. That name does not exist —_iroh4is a misreading of the DNS wire format, where\x05_iroh(label) followed by\x34(52 = the length of the z32 label, and ASCII4) looks like the text_iroh4. The real record (_iroh.<z32>.dns.iroh.link) resolves normally:
dig TXT _iroh.57mr4k9n….dns.iroh.link @1.1.1.1
# status: NOERROR TXT "relay=https://use1-1.relay.n0.iroh-canary.iroh.link./"Use
./omp-phone.sh diagto see both lookups for the endpoint this deployment uses.
Bundles under dist/ and harness-web/ are vendored verbatim from their upstream projects and
remain under their own licenses:
| Path | Upstream | License |
|---|---|---|
dist/ |
@oh-my-pi/collab-web (can1357/oh-my-pi) |
MIT |
harness-web/ |
giuliastro/harness-remote | Apache-2.0 |
collab-serve.ts, iroh-serve.sh, phone-link.sh, omp-phone.sh, install.sh, the systemd
units and this document are this repo's own code (MIT).