hardware
Hardware
The Hardware page (/hardware) is the host- and physical-level infrastructure view
— the section no customer ever sees. It acts on the VM itself, beneath every workload.
It is the most locked-down page in god mode: it requires hardware.manage, which is
super-admin only and can never be granted to a sub-admin (see the
privilege catalog). It is also god
launch-mode only — switch to Beta or Production and Hardware disappears from the nav
entirely. Without the privilege the page shows a locked state.
The page itself carries a god-mode notice restating both facts: it's visible only in god launch mode, and its controls reach the physical/host layer — reboot, recreate, and key rotation operate on the machine, not the apps running on it.
Host inventory (real)
A KPI row summarizes the fleet at the host level:
| Stat | Meaning |
|---|---|
| Hosts | Every visible machine, split into provisioned, connected, and provider-only origins |
| Raw disk | Total provisioned disk across all VMs |
| Stale agents | Agents behind the current version (accented — they need an upgrade) |
| On overlay (WG) | Hosts that joined the WireGuard mesh |
The Host inventory table lists every physical/virtual machine with host-level facts the customer view never exposes: name, public IP and WireGuard IP, provider, region, raw disk, live CPU and memory meters, agent version, and status. This inventory is real — it comes from the same fleet source the rest of god mode uses (self-registered agents that activated and heartbeat in), so a freshly installed box shows up here.
Node telemetry — the full picture without SSHing in
Each node row has a collapsible telemetry panel (reported by the agent's heartbeat), so you can read a machine's real state from the dashboard instead of opening a shell:
- GPUs — model, VRAM used/total, and utilization for each GPU (GPU nodes only).
- Software stack — OS, Docker version, and — on a GPU node — the NVIDIA driver, CUDA, and nvidia-container-toolkit versions.
- Containers — the running-container inventory (name · image · state), so you see exactly what a node is running.
This updates every heartbeat, so an idle box, a GPU under load, or a container that just exited all show here within a minute — no SSH required. (See GPU workloads for GPU nodes.)
Danger zone — prototype (not wired)
The Danger zone presents three irreversible, host-level operations, each behind a confirmation drawer that spells out the impact:
| Action | Intended effect |
|---|---|
| Reboot VM | Hard-restart the host; workloads bounce, the IP is preserved (~30–90s downtime) |
| Recreate VM | Destructive — wipe and reprovision from a fresh OS image, then restore from the latest .hsi; a new IP is assigned and DNS flipped |
| Rotate host keys | Re-issue the host SSH key pair; existing fingerprints are invalidated and sessions drop |
These actions are a prototype and do not execute. Host-level operations run on the agent or the provider API, which isn't wired yet. Confirming any of the three shows a "coming soon" toast and changes nothing — the UI states this honestly and never claims a destructive op happened. (An earlier version of this control incorrectly flipped launch mode; it no longer mutates any state.) Treat the danger zone as a preview of the surface, not a working control.
The inventory above is real; only the danger-zone actions are stubbed.
Related
- God mode (admin overview) — the RBAC gate, the super-admin-only rule, and launch-mode gating
- Identity & access — why
hardware.manageis super-admin only - Fleet — the customer-safe, workload-level view of the same machines
- Provisioning — spinning up and destroying whole servers