# What an always-on agent machine costs

> What an agent machine costs by the hour and month at shedix's prices, what leaving it on adds up to, and how auto-sleep, stopping and deleting change it.

*2026-11-01 · costs, agents, remote-dev*

Leave a machine on all month and it runs about 730 hours: 24 hours times 365 days, divided by 12. Multiply that by the hourly price and you have the cost of never turning it off. For a machine that only works when you hand it a task, most of those hours are idle.

## By the hour and by the month

shedix prices each size by the hour while it runs. Here is what that means if a machine never stops:

| Size | Machine | Disk included | Per hour | 730 hours |
|---|---|---|---|---|
| S | 2 vCPU, 4 GB RAM | 30 GB | $0.06/hr | $43.80 |
| M | 4 vCPU, 8 GB RAM | 45 GB | $0.16/hr | $116.80 |
| L | 8 vCPU, 16 GB RAM | 60 GB | $0.32/hr | $233.60 |
| XL | 16 vCPU, 32 GB RAM | 75 GB | $0.64/hr | $467.20 |

XL is available from the API and CLI. Custom sizes go from 2 to 16 vCPU and cost $0.032 per vCPU-hour plus $0.004 per GB of RAM per hour, so a 6 vCPU machine with 24 GB is 6 × $0.032 + 24 × $0.004 = $0.288/hr.

The price covers the machine. Your agent's model usage is separate: you sign in to Claude Code or Codex yourself on the machine, so that usage stays on your own Claude or ChatGPT account.

## What real use looks like

Here is what a few common patterns come to:

- **Working hours:** 8 hours on each of 22 workdays is 176 hours. That's $10.56 on an S, $28.16 on an M and $56.32 on an L.
- **A few hours a day:** 4 hours on 22 workdays is 88 hours, or $5.28 on an S and $14.08 on an M.
- **One long run:** a 6-hour agent job on an M is $0.96.
- **Never stopped:** an M left on all month is $116.80.

An M that runs during your working hours costs about a quarter of one that never stops. The difference is idle time, and cutting it is what auto-sleep is for.

## Idle detection and auto-sleep

shedix watches each machine for signs of use and stops it after 60 minutes without any. A machine counts as in use while any of these is true:

- an SSH session is connected
- a terminal or VS Code tab is open
- an agent is writing its conversation
- the CPU is busier than a quarter of a core
- the network moves more than 1 MB in 30 seconds

After an hour with none of that, the machine stops by itself. Its files stay on the disk. Running processes end.

For cost, this caps what forgetting a machine can do. A forgotten M costs about $0.16 for the idle hour before it stops, instead of $7.68 over a 48-hour weekend.

Two habits keep auto-sleep working for you:

- **Close what you're not using.** An open terminal or VS Code tab counts as use for as long as it stays open, and so does a connected SSH session.
- **Use Keep awake on purpose.** A job in tmux with nobody connected counts only while it keeps the CPU or the network busy. For long unattended work, turn on Keep awake for that machine, and turn it off when the work is done. An L kept awake over a weekend is 48 × $0.32 = $15.36.

## Stopping versus deleting

Sleep is a full stop, the same as pressing Stop yourself. Stopping and deleting both end a machine's hours, in different ways:

- **Stop:** no hourly charge while it's stopped. The disk and everything on it stay: your repositories, installed dependencies, Claude Code's sign-in and its conversations. Starting it again takes about 45 to 60 seconds, and it comes back with a new IP address. Processes don't come back on their own, so restart what you need; for Claude Code, run `claude --continue` in the project folder.
- **Delete:** the machine and its disk are gone for good. A new machine is ready in about 60 to 100 seconds, starting clean.

Stop a machine you'll come back to. Delete one whose work is merged and pushed, or one that held something you don't want left on a disk. From a script or an agent, both are one call:

```bash
curl -sS -X POST "$SHEDIX_URL/api/machines/$NAME/stop" -H "Authorization: Bearer $SHEDIX_TOKEN"
curl -sS -X DELETE "$SHEDIX_URL/api/machines/$NAME" -H "Authorization: Bearer $SHEDIX_TOKEN"
```

## Disk costs

Each size includes the disk in the table above. Disk beyond that is $0.10 per GB a month, so 30 GB extra is $3.00 a month. Extra disk is priced by the month, not by the running hour.

That mirrors the cloud underneath. Google Cloud stops charging for a stopped instance's CPU and memory, but [its disks are charged for as long as they exist](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), whatever state the instance is in, and its docs on [stopping an instance](https://docs.cloud.google.com/compute/docs/instances/stop-start-instance) say the same about disks and external IP addresses.

To see what's using space before you ask for more, check from the machine's terminal:

```bash
df -h /
du -sh ~/work/* | sort -h
```

Old Docker images, package caches and forgotten worktrees are worth checking first.

## A rule of thumb

Pick the smallest size that keeps your agents from waiting on the machine. Let auto-sleep catch what you forget. Stop machines you'll return to, delete the rest, and save Keep awake for the jobs that need it. For what a single test run costs on each size, see [our post on running tests on a bigger machine](/blog/run-tests-on-a-bigger-machine).

Every size is on the [pricing page](/pricing), and the [FAQ](/faq) covers sleep and stopped machines. When you're ready, [start a machine](/).