Skip to content

spinloop metrics

What every engine you run is doing with its hardware — resource use, token and request counters, and what the node it runs on is.

spinloop metrics                      # the fleet.yaml in this directory
spinloop metrics --fleet ./cluster.yaml
spinloop metrics --env prod           # one registered environment, no file needed
spinloop metrics -w                   # redraw in place every 60 seconds
spinloop metrics --cost               # add what each priceable node has spent

Which target

The three ways every command that reads engines takes one:

target
--env <name> one registered environment
--fleet <path> (-f) that fleet file
neither the fleet.yaml in the working directory

--env and --fleet name two different things, so passing both fails saying so. With none of the three resolvable the command fails naming all of them.

Formats

--format selects one of four. All four report the same facts; they differ in how much of the screen they spend.

format what it draws
gauge (default) a header line per node, then the current reading as gauges
bar the same, with each series as a sparkline of the retained history
table every fact on its own line, then the counters and figures
json one object per node, for a program to read

table is the one to use when you want everything a node can say:

node:         prod
state:        running
instance:     i-0abc123
instanceType: g6e.xlarge
runner:       llamacpp
model:        org/qwen3-27b
version:      1.41.0
endpoint:     http://198.51.100.1:8000/v1
uptime:       2h 15m 0s
active:       2m 5s ago

  running:          1
  prompt tokens:    12000
  …

Which flags apply to which nodes

Some facts only one kind of node has. A flag that asks for one applies to the nodes that can answer it and leaves the rest exactly as they read without it — a fleet of daemons asked for a cost shows no cost and still succeeds.

flag / field cloud environment daemon node
--cost priced from its instance type and region —
instance, instanceType reported —
version reported —
endpoint reported —
everything else reported reported

A daemon node runs on a machine you already own: it has no hourly rate, no instance id, and no address the control plane publishes. Its release is on its status row instead.

Watching

-w/--watch redraws every 60 seconds, clearing the screen each time rather than accumulating scrollback, and exits cleanly on interrupt. For an interactive view that also drives the nodes, use spinloop dashboard.

See also