Self-hosted GitHub Actions runners

Your CI, on your cloud. Without running a single runner yourself.

Commit an airship.yaml, point runs-on at an airship- label, and every job gets its own machine in your cloud account. Airship launches it, watches it and tears it down.

AWS · Google Cloud · Azure

Start for free Read the docs

acme/Runners

live

Busy14

Idle pool2 / 2

Spot71%

Cap14 / 40

JobLabelMachineRegionStateTime

test (ubuntu, node 24)acme/webairship-buildAWSc7i.xlargespotus-east-1busy2m 41s

train-smokeacme/modelsairship-gpuGCPg2-standard-8spotus-central1-blaunching0m 12s

e2e (shard 3/4)acme/webairship-buildAWSm7i.xlargeus-east-1running0m 38s

msbuild releaseacme/desktopairship-windowsAzureStandard_D4s_v5eastusbusy6m 02s

lintacme/apiairship-smallAWSt4g.mediumspotus-west-2pending—

cargo build --releaseacme/engineairship-armAWSc7g.2xlargespoteu-west-1finished4m 19s

.github/workflows/ci.yml

jobs:
  test:
    runs-on: airship-build--${{ github.run_id }}
    steps:
      - uses: actions/checkout@v5
      - run: make test

Your account, your bill

Machines launch in your AWS, Google Cloud or Azure account. You pay the cloud, at your rates.

One file per repo

airship.yaml says what a runner is: machine type, image, region, spot, GPU. runs-on picks it.

Fresh machine per job

Every job gets a new machine that registers for that one job and is gone when it ends.

Spot, then fall back

List instance types in order. Spot capacity first, on-demand when the cloud says no.

Features

Everything a runner needs, in one file

Each repository describes its runners in airship.yaml. Airship turns that into machines.

Three clouds, one formatThe same airship.yaml launches EC2 instances, Compute Engine VMs or Azure VMs. Change cloud: and nothing else.Spot first, on-demand when it runs outList instance types, regions and spot flags in order. Airship walks the list until a cloud says yes, and records why the others said no.Warm runners, on a scheduleKeep a few machines booted so jobs start in seconds: always, in office hours, or for a while after work arrives. Per repository or for the whole organization.GPU runnersAdd gpu: and Airship attaches the accelerator and installs the NVIDIA driver on Ubuntu images that lack one. T4, L4, A100, H100 and up.Windows, ARM and your own imagesUbuntu by default; any AMI, Compute Engine image or Azure gallery image of your own. Windows images get a Windows boot script.Cache in your own bucketOn AWS, cache: true puts a proxy on the runner that serves actions/cache from an S3 bucket in your account, in the runner's region. No size limit but yours.A developer APIActivate repositories, connect clouds and change access rules from scripts or your own tooling, with a scoped API token.

airship.yaml

runners:
  - name: build
    cloud: aws                 # or gcp · azure
    instance_type: c7i.xlarge
    region: us-east-1
    labels: [airship-build]
 
  - name: windows
    cloud: azure
    instance_type: Standard_D4s_v5
    machine_image:
      publisher: MicrosoftWindowsServer
      offer: WindowsServer
      sku: 2022-datacenter-g2
    labels: [airship-windows]

airship.yaml

runners:
  - name: gpu
    cloud: gcp
    instance_type: n1-standard-8
    region: us-central1        # any zone with capacity
    gpu: nvidia-tesla-t4
    preemptible: true
    labels: [airship-gpu]

$ nvidia-smi --query-gpu=name,memory.total

name, memory.total [MiB]

Tesla T4, 15360 MiB

terminal

$ curl -H "Authorization: Bearer $AIRSHIP_TOKEN" \
    https://api.airshipci.com/api/v1/repo \
    -d '{"repository": "acme/web", "active": true}'
 
{"fullName": "acme/web", "activated": true, ...}

How it works

We run the fleet. You own the machines.

Airship's control plane holds your sealed cloud key, listens to GitHub, and launches, watches and terminates machines in your account. Nothing long-lived runs on your side.

Connect a cloud

GitHubAirshipYour cloud

job queued: airship-build

reads airship.yaml

RunInstances c7i.xlarge · spot

runner registers for this one job

job runs

shuts itself down

A job asks. A machine answers.

GitHub tells Airship a job wants an airship- label. Airship finds the runner in that repository's airship.yaml and launches it in your account, registered for exactly that job.

Labels without the airship- prefix cost no API calls.

cloudaws · us-east-1

access_keyAKIA••••••••••••

secret_key••••••••••••••••••••

sealed · bound to this row

opened forRunInstances→ dropped

Your key, sealed at rest

The access key you connect is encrypted, tied to its own database row, and opened only for the length of one cloud call. It is never logged and never sent back.

Give it a role scoped to what Airship launches.

pending
launching
running
busy
finished

finished, and why

completed
cancelled
idle
launch_failed
lost
max_lifetime

Nothing left running

Every runner moves through one lifecycle, and every ending is recorded. A reconciler compares the cloud with the database and removes machines nobody is using, so a lost answer from the cloud doesn't become a bill.

Idle 10 minutes, or 8 hours old: the machine goes.

AWSEC2 · S3 cacheconnected
Google CloudCompute Engine · GPUsconnected
AzureVirtual Machinesconnected

runs

ubuntu2404
Windows Server
arm64
nvidia-l4
spot
custom AMI
gallery image

Runs where your account already is

Connect one cloud or all three. Each runner names its cloud, and machines launch with your credentials, in your regions, on your bill.

Change cloud: in airship.yaml and nothing else.

Security

Your jobs run on machines you own

Airship reads airship.yaml and job events from GitHub and makes cloud calls with your key. Your code, secrets and build output stay between GitHub and your machines.

SEALED KEYSEncrypted at rest, opened per call
YOUR NETWORKYour network, your subnets
ONE JOBOne machine per job, then gone
SCOPED CACHEShort-lived grants to one bucket
SELF-HOSTRun the control plane yourself

Runners register with a just-in-time config for a single job, so a machine can't pick up work it wasn't launched for. Only labels starting with airship- are looked up, and only in repositories someone activated.

Get started

Your first runner in four steps.

1Install the GitHub App
2Connect a cloud
3Activate a repository
4Commit airship.yaml

Start for free

airship.yaml

runners:
  - name: build
    cloud: aws
    instance_type: [c7i.xlarge, m7i.xlarge]
    preemptible: [true, false]
    labels: [airship-build]
    extra_config:
      cache: true