Add a second GPU host

Coming soon.

Not tested, not supported yet

Multi-host is in the architecture from the beginning. The control plane holds no GPU state, the scheduler places sessions across hosts, and the admin UI is built around a fleet rather than one machine. None of that is speculative.

What does not exist yet is a tested way to install the second machine. There is no agent-only Compose file, no installer, and no verification run behind it. Any recipe published today would be guesswork.

What already works

Each host runs one node agent. It reports its GPUs, capacity and readiness to the control plane, and receives sessions from it. The pieces are in place; the packaging and the testing are not.

Hosts can use different local storage roots. If you want saves to follow a user between machines, back those roots with shared storage. Otherwise set the locality placement policy so Quasar prefers the host that already holds that user’s home directory.

If you want to help

The repository has a multi-agent Compose overlay used for scheduler testing. It runs several agents on one machine and is not a second-host installer.

If you want to work on real agent-only packaging, start with GPU and remote workflows. The thing to get right is the runtime contract the base node-agent service carries: host networking, GPU and /dev/uinput access, the PulseAudio sidecar, the managed-home root mounted at the same path inside and outside the container, and the enrollment flow that issues a per-node secret on first contact. Copying five environment variables across is not enough, which is most of why this is not shipped.