spark
loading…
load
gpu —
compute busy
cpu —
1-minute load
memory —
unified, ram + gpu
gpu temp —
degrees c
who's on it
per-user charts →hardware
who's using it.
loading…
right now
cpu
cores in use, stacked per person.
memory
ram + gpu memory, stacked. they come out of the same unified pool on the spark.
gpu
compute
share of the gpu's streaming multiprocessors each person's processes keep busy.
gpu memory
cuda allocations. docker containers count toward the owner.
by day
cpu hours
core-hours per day. one core flat out for an hour is 1.
gpu hours
hours of full gpu compute per day, summed from busy time.
totals
processes
logins
the guard stops a guest's biggest gpu process when their ram + gpu goes over the cap. cpu and ram caps are enforced by the kernel.
claude accounts
add one claude provider instance per account and point its claude_config_dir path at the folder shown here.
services
nightly updates
checked every 30 minutes. if a new nightly is out while an agent session is running, the update waits until the session ends.
vibeproxy
accounts
open sign-in ↗
point your tools at http://127.0.0.1:8317 with the key in ~/.config/cliproxyapi/client-api.key. add -thinking-8000 to a claude model name for an 8k thinking budget (up to 31999), or use gpt-5-high-fast style names for codex.
config
show config
tailscale
peers
serve config
logs
pick a service and press load.