Skip to content
Spinbox is experimental: APIs change and data may be lost. Status lists what is built and what proves it.

Profiles

View Markdown

The module’s defaults are a small installation: at most one runner (max_runners), of 32 to 48 vCPUs (runner_vcpus), on spot, given back once nothing has run for runner_idle_minutes. Each profile below says who it is for, what to change from those defaults and why, what that trades, and what has not been tried. Cost says what each part is billed at; this page does not repeat it.

The settings are arguments of module "spin", in your copy of examples/complete (its main.tf, which already sets runner_idle_minutes, quiet_hours, time_zone and max_runners). Its terraform.tfvars takes only the example’s own variables, such as spin_version and domain. Every argument, with its default, is in the module reference.

Two things no profile changes yet. How many workspaces fit on one runner has not been measured, so every max_runners below is a starting point to correct against what you see. And the control plane and the proxy run all the time: there is no standby that stops them while nobody works, nor a smaller installation for one person than the one the module makes.

A model, not a bill: us-east-1, AWS’s prices of 2026-11-19, spot runners at their 30-day median price (Spot), the part that runs all the time at the one measurement so far (With nobody working), and, the largest unknown, eight workspaces on a runner. Each line is in dollars a month.

ProfileWhat the model assumesAboutEach person
One personOne or two workspaces through the working day$165$165
One person with agentsThe person, and four agents’ workspaces through the working day$230$230
Agents around the clockThe person, and three agents’ workspaces never stopping$560$560
Small teamFive people$210$42
Team22 people in two time zones, a few agents, light CI$950$43
Large100 people in three time zones, agents, CI and prebuilds, a larger control plane$4,100$41

The share of a person moves most with density: at four workspaces a runner rather than eight, the team of 22 is about $1,560 a month, and at sixteen about $650. With every runner on demand, each line is more than double. What a workspace writes to the bucket, and what it sends out of the region, are guesses too; they matter most for agents.

For a developer on their own, with a workspace or two open while they work.

The defaults fit. What is worth setting:

module "spin" {
# ...
quiet_hours = "20:00-07:00"
time_zone = "Europe/Madrid"
monthly_budget_usd = 200
budget_emails = ["you@example.com"]
}
  • quiet_hours and time_zone: at night the runner goes as soon as nothing runs, instead of after runner_idle_minutes.
  • monthly_budget_usd and budget_emails: a mail when the month passes 80% or 100% of it, or is forecast to. It stops nothing.

Trades. The first workspace after the runner has gone waits while a new one starts (Waiting for a host). A lower runner_idle_minutes saves runner hours and makes that wait more frequent.

A smaller runner. Most of one person’s bill is a runner of 32 vCPUs or more kept up for a few cores of work, and the parts that run all the time. A runner of 8 vCPUs would cut the first:

runner_vcpus = { min = 8, max = 8 }

The module accepts it, and the group then asks for the 2xlarge sizes of the default families with local NVMe. In the model above, it takes one person from about $165 to about $100 a month. No installation has run runners that small. A preset as large as the runner leaves nothing for the host itself, so keep presets well under it (Installation settings); and with max_runners still one, a workspace that does not fit waits for room.

For a person who runs agents in workspaces of their own beside the one they work in, including workspaces that run for days.

module "spin" {
# ...
max_runners = 2
runner_rollout = "rolling"
monthly_budget_usd = 600
budget_emails = ["you@example.com"]
}
  • max_runners = 2: if the agents’ workspaces do not fit on one runner, the next one can go to another runner instead of waiting for room on the first. The control plane adds a runner only when a workspace has waited ten minutes (Scaling from zero).
  • runner_rollout = "rolling": with the default, idle, a new release reaches the runners only once the fleet empties, which a fleet with a workspace that never stops does not. Rolling replaces them at an update, each workspace suspended and resumed on the new host (Update).
  • A budget: a workspace that runs for days keeps a runner up for days, and the bill shows it.

Not Terraform, and the setting a workspace that runs for days meets first: the installation suspends a workspace that has run as long as it allows, used or not, 12 hours by default. An administrator sets it, from no cap to a week:

Terminal window
spin admin settings get workspace_max_running_minutes
spin admin settings set workspace_max_running_minutes 0

Zero caps none. The owner of one workspace can instead keep it awake for a while (Workspaces).

Trades. Without the cap, a forgotten workspace keeps its runner up until somebody stops it. On spot, AWS can take a runner back at any time: its workspaces are suspended to the bucket and resumed on another, which an agent in the middle of a task sees as a pause. runner_spot = false avoids that at the on-demand price (Spot). A workspace that runs for a long time without stopping also meets the disk chain’s limit.

Not tried. The agent features are built and tested, and the least used day to day. Agents running for days on one installation, and how often spot interrupts a runner over a week, have not been measured.

For about five people in one time zone.

module "spin" {
# ...
max_runners = 2
quiet_hours = "20:00-07:00"
time_zone = "America/New_York"
monthly_budget_usd = 400
budget_emails = ["ops@example.com"]
}
  • max_runners = 2: at the density assumed above, five people fit on one runner; another covers being wrong about it, and busy days.
  • quiet_hours in the team’s time zone: the runner goes as soon as the last person stops.

Trades. The next runner starts only after a workspace has waited ten minutes, so on a busy morning someone waits. Once every runner is up, a workspace that does not fit waits for room.

Not tried. How many workspaces fit on a runner. The bill shows what each workspace held, and the runners’ capacity nobody held, which is the measure to correct max_runners by.

For twenty to twenty-five people, in one or two time zones.

module "spin" {
# ...
max_runners = 6
quiet_hours = "22:00-06:00"
time_zone = "Europe/Berlin"
monthly_budget_usd = 1500
budget_emails = ["ops@example.com"]
}
  • max_runners = 6: what the model needs at its busiest hour at eight workspaces a runner. At four it would need eleven, at sixteen three. request_quota (on by default) asks AWS for the spot vCPUs that many runners of runner_vcpus’ maximum take (Before you start).
  • quiet_hours only over the hours neither time zone works, or none: inside the window a gap between two people empties the fleet, and the next one waits for a runner.
  • Keep runner_instance_types broad: each type and zone it matches is one more spot pool to start from (Spot).

Trades. Runners are given back one at a time, once a runner has held nothing only it holds for runner_idle_minutes; a lower value gives runners back sooner and starts them more often.

Not tried. Density, as above, and twenty people on one installation.

For about a hundred people across time zones, with agents, CI jobs in workspaces, and prebuilds.

module "spin" {
# ...
max_runners = 20
runner_rollout = "rolling"
instance_type = "m8i.xlarge"
controlplane_volume_gb = 200
monthly_budget_usd = 6000
budget_emails = ["ops@example.com"]
}
  • max_runners = 20: the model’s busiest hour at eight workspaces a runner; between ten and forty over the densities it was run at.
  • runner_rollout = "rolling": a fleet working across time zones may never empty.
  • instance_type and controlplane_volume_gb: the control plane also runs the database and the stores of logs, traces and metrics, and nothing in the module sizes it by how many people use it. These two values are a guess, not a measurement: watch its machine (Observability) and size it from that. A change replaces the machine, as an update does.
  • Without quiet hours: someone is always working.

Trades. Spot capacity for many runners at once is less sure than for one: AWS’s placement score for the default types fell at 20 runners in us-east-1 and held in us-west-2 on the day it was read (Spot). A host class can fall back to on demand when spot is refused (Admin → Runners). The control plane and the proxy are still one machine each (Limitations). Prebuilds keep their disks in the bucket until retired, which an administrator can make automatic for unused ones in the installation’s settings.

Not tried. A hundred people, the control plane at that load, the proxy (proxy_instance_type) at that traffic, and twenty runners on spot at once.

  • After the installation has run for a while, switch on its cost allocation tags in bootstrap, or the Costs page stays empty (Reading the bill).
  • runner_spot, runner_instance_types and runner_vcpus change what a runner is; keep the types to one CPU generation, so that a suspended workspace can resume on any of them rather than boot cold.
  • How long a workspace may sit unused before it is suspended is an installation setting, like the running cap above: Installation settings.

Installations run Spinbox v20261027.01; the reference is generated from spin v20261114.01 (51f7e0dea). This page as Markdown, for a model or a terminal.