# Runner providers

A runner provider is a cloud account where SuperCI starts machines for your jobs. Add them under **Runners**. You can use one or several; with several, you set their [order](/docs/order-and-limits).

| | AWS | Cloudflare | Modal |
|---|---|---|---|
| A job runs in | an EC2 machine | a container | a sandbox |
| Largest machine | any size EC2 has | 4 CPUs, 12 GB, 20 GB disk | 64 CPUs, 336 GB |
| Systems | Linux and Windows, x64 and arm64 | Linux, x64 | Linux, x64 |
| GPUs | T4, L4, A10G, L40S | none | any Modal offers |
| Price | spot or on-demand, as AWS bills it | CPU while it is used, memory while it runs | per second, for the exact size |

## AWS

Machines are spot by default, with on-demand as a provider of its own (see [Spot machines](/docs/spot-machines)).

- **Regions:** choose one or several. Spot machines start where a job costs least, across all of them, the first region counted 15% cheaper (its cache is there); a cheap machine in the next region comes before a dear one in the first. On-demand machines start in the first, then the next.
- **What a job finds:** a machine image with the tools of GitHub's own Ubuntu 24.04 runners, published by [RunsOn](https://runs-on.com) as public images. Docker is there.
- **What SuperCI creates:** a role the control plane assumes with short-lived credentials (no stored keys), and a network for jobs in each region with no way in. Both are tagged `superci-plane`, and removed when you remove the provider.
- **Your account's limits:** AWS gives new accounts a small spot quota. The dashboard shows it for each region, with a link to ask AWS for more.

You can also start machines in [your own network](/docs/private-networks).

## Cloudflare

Jobs run in Cloudflare Containers, sized per job. They start fast and cost nothing while no job runs. Containers need Cloudflare's Workers Paid plan.

- **Location:** choose the area containers start in.
- **What a job finds:** a small Ubuntu image with GitHub's runner. Docker works inside a job, with one difference: containers share the machine's network, so two services cannot use the same port.
- **What SuperCI creates:** a launcher (a Worker) and its container application, named `superci-runners-<id>`: it starts a container for each job your control plane sends it.

### GitHub's full image (experimental)

The small image has GitHub's runner and little else. To run jobs inside GitHub's own Ubuntu 24.04 runner image, with everything it has, set its address under **Runners → Cloudflare → GitHub's full image**:

```text
https://images.superci.dev/ubuntu-24.04-full/de14111c9565405e3d9581170f3c63ba
```

The image is far larger than a container's disk, so it is not copied there: a job fetches only the parts it reads, and each part is checked against the address before it is used. A job whose image cannot be loaded fails saying so.

## Modal

Jobs run in Modal sandboxes, up to 64 CPUs, with any GPU Modal offers.

- **What a job finds:** a container with GitHub's runner. There is no Docker inside a sandbox, so jobs that build or run containers belong on AWS or Cloudflare, and GitLab jobs (which run in Docker) do not run on Modal.
- **CPUs** are Modal's cores, two of a job's CPUs each: a job asking for 4 CPUs gets 2 cores, and `nproc` says 2.
- **GPUs** need a payment method on your Modal workspace.
- **An image of your own:** a job can start from one with your tools in it already. Make it `FROM ghcr.io/actions/actions-runner`, name it once (`superci runner_images create rust-ci --reference=ghcr.io/acme/ci:2026-10`, with `--modal-secret=NAME` for a private registry), and say it in the label: `superci/image=rust-ci`.
- **Functions:** `modal-functions` is a row of its own in the order, like AWS on-demand: preemptible Modal Functions at a third of a sandbox's price, taken out until you put them in (`superci runners order modal-functions modal`). A job Modal preempts stops within seconds, and runs again by itself further down the order once its run has finished.
- **What SuperCI creates:** a launcher, an app named `superci-runners-<id>`: it starts a sandbox for each job your control plane sends it.

## Removing a provider

**Runners → the provider's settings → Remove** stops jobs going there, then deletes what SuperCI made in that cloud, with your sign-in there. The dialog lists what goes before you confirm.
