Remote Orchestration

[ForbiddenEra]

Problem / Use case

I'd like to be able to start ralph-tui with only remote instances; right now adding a remote adds a tab for that remote - I just want the first 'local' tab to be gone and not try to do anything locally.

Proposed solution

Alternatives considered

No response

Additional context

No response

[subsy]

@ForbiddenEra this is difficult with the current architecture as the TUI is dependent on a local engine being available. But I'll look into it. Stay tuned

[ForbiddenEra]

@ForbiddenEra this is difficult with the current architecture as the TUI is dependent on a local engine being available. But I'll look into it. Stay tuned

k I mean we can add tabs as remote so - what does it depend on local? I'm not necesarily against installing it locally still but I want to manage remote ones.. I have diff machines with diff gpus, would be nice to command them all from one spot even if I'm not actively working on that machine at the time.. really I just envisioned it working the same just no first tab when a remote is enabled and no local or something..?

[benkauffman]

Following this — I ran into exactly the same problem building ctrl, which is a visualization layer for coding agents (not an orchestrator, but related problem space).

The approach I landed on: a relay. Each daemon registers itself to a relay server when it starts. The client connects to the relay and the relay brokers the WebSocket connection between them. Neither side needs a reachable IP or open port. You get the same UX as a local connection — it just happens to go through the relay when the daemon is on a remote machine.

This wouldn't solve the "local engine required" problem directly for ralph-tui (you'd still need the engine running on the remote machine), but it might be worth pulling out the transport layer as a separate concern. If the TUI could talk to a remote engine over a relay, you could manage all your GPU machines from one TUI instance without exposing anything.

Not trying to push ctrl here — it's a different tool. Just sharing the architecture pattern in case it's useful. Happy to dig into the relay implementation details if that helps.