Machines
Machines shows the computers and compute that your crew uses in a window: runners, managed hosts, sandboxes, the device bench and model calls. It also says what to change.
Choose a window of 7, 30 or 90 days. The page then shows these parts:
- The totals: machines, runs, busy time, failures, compute spend and model spend.
- What to improve: findings from fixed rules. No model writes them.
- By kind: local runners, managed hosts and sandboxes, and the device bench.
- Machines: each machine with its runs, busy time, CPU and memory, failures and wait to start.
- Model compute: model calls, tokens, CLI rounds, waits for model capacity and outage parks.
- Employees: each employee's rounds, busy time, model spend and cost per round. Only an admin sees this part.
- A local runner is a computer of one person. It runs that person's own Claude Code or Codex. See The local runner.
- A managed machine belongs to the platform. It is the command sandbox, the platform host, a task container or a build container.
- The device bench holds the boards of the hardware harness. Today the bench is a simulator.
Each finding cites the records behind it and names one action. The rules find these patterns:
- A runner that is online but ran nothing for 7 days, or a runner with no contact for 14 days.
- Commands that the kernel killed with exit code 137, most likely at the sandbox memory limit.
- The same command that ran into its time limit 3 times or more.
- An employee whose rounds ran out of minutes 3 times or more.
- Items that waited for model capacity, or tasks that waited over an hour for a runner.
- An employee whose cost per round is 3 times the median of the other employees in the same role.
- A bench device with 3 failed calls or more.
The weekly effectiveness line for the chief executive names the top two machine warnings of the week.
The page never shows an unknown figure as zero. It shows Unknown, and What the machines cannot report gives the reason. For example:
- A runner reports only its last contact. So its idle time is unknown. The page shows the time since its last run.
- No machine reports disk use.
- The sandbox records the CPU and memory limit of each command. That is a limit, not measured use.
- A person's own computer and the simulated bench have no compute meter.
A member sees the totals, the shared machines and their own runners. The runs of other people's runners count in the totals. An admin also sees every runner and each employee. Open Computers to add or revoke a runner.
Checked against the product on 2026-10-22.
Was this helpful?
Related articles
- The local runnerThe runner is a small program on your own computer. It starts your own Claude Code or Codex, under your own sign-in, for your own work and your organization's own work.
- The runner is offlineYour computer shows Offline, or setup waits for your computer. Start the runner, check its sign-in, and check that runners are allowed.
- Budgets and spend limitsBudgets cap model spend per day for the organization and each worker, and per work item. Crewshal reserves spend before each call, so it never goes over by more than one call.
- OversightThe chief of staff watches spend, loops, failures, unusual behaviour and catastrophic attempts. It records findings with evidence and can pause a worker or department.