machine.dev Docs

What is machine.dev?

machine.dev gives you on-demand GPU and CPU runners for GitHub Actions. Add runs-on: machine/gpu=t4g to a workflow and your job runs on a 16 GB GPU. Pricing is per-minute in US dollars: spot from $0.00261/min (T4G ARM), on-demand from $0.01400/min.

What you get

• 3 lines of YAML, no infrastructure to manage. Add machine to runs-on: and that’s it.
• 6 NVIDIA GPU types (T4G, T4, L4, A10G, L40S, RTX 6000) plus AWS Trainium and Inferentia2 accelerators.
• CPU runners from 2 to 64 vCPUs in X64 (Intel/AMD) and ARM64 (Graviton) flavors.
• Per-minute billing in real dollars. No credits. Cold start ~1 minute. You only pay for runtime.
• Configurable storage. Every runner gets a 100 GB gp3 EBS volume by default, expandable to 16 TB with custom IOPS and throughput.
• Instance metrics collected by default and rendered as sparkline charts on the dashboard.
• Spot pricing saves 70-90% versus on-demand. Combine with checkpointing for resilient long jobs.

Quickstart

1. Sign up for a machine.dev account. New accounts get $10 of free compute.

2. Install the Machine Provisioner on your account or organization.

3. Enable self-hosted runners for the repos you want machine.dev to run jobs for.

4. Add this to a workflow file (e.g. .github/workflows/test.yml):

   jobs:
     gpu-job:
       runs-on: machine/gpu=t4g/tenancy=spot
       steps:
         - uses: actions/checkout@v4
         - run: nvidia-smi

5. Push the workflow. Watch the job pick up a machine.dev runner in your Actions tab.

This minimal demo runs in about a minute on a T4G spot runner. T4G spot starts at $0.00261/min.

For a full walkthrough see the Quickstart guide.

How machine.dev fits with GitHub Actions

When you add runs-on: machine/... to a job, GitHub Actions hands the job to the Machine Provisioner. We launch a fresh AWS instance matching your labels, the GitHub Actions Runner agent inside the VM dials home and registers itself as an ephemeral self-hosted runner attached to your org or repo, GitHub dispatches the job, and the runner shuts down when the job finishes. You only pay for the seconds the runner was up.

Free tier

Non-paying teams can run one job at a time, with a 30-minute limit per job. Buying compute or subscribing removes the 30-minute limit. Promotional grants alone do not count as a purchase. See Billing & limits for paid concurrency limits.

New accounts get $10 of free compute to try machine.dev. That is ~63 hours of T4G GPU time at today’s spot rate (from $0.00261/min), or several thousand minutes of CPU time. Sign up and add a workflow.

Where to go next

• Quickstart: first workflow in 5 minutes
• GPU Runners: every GPU type, configuration, and price
• CPU Runners: X64 and ARM64 from 2 to 64 vCPUs
• Pricing: full per-minute rates in US dollars
• CPU vs GPU: pick the right runner for your workload
• Use Cases: real workflows you can fork
• Cost Optimization: spot, checkpointing, right-sizing