CI-Ninja runners
Make GitHub Actions faster without rewriting your workflows.
Your CI shouldn't be the thing slowing down your engineers. Run the workflows you already have on faster machines, measure the difference on your own repository, and switch when the numbers say so.
No charge until they run out. To keep going you add a card and set a monthly spending cap. Benchmark mode is free, with no card.
Measured, not promised
Up to 2× faster. Here is how we know.
Speed-ups depend on your jobs. These are the numbers we have, with how they were measured.
One customer's real run
45% less wall time (1.8×)
- Workload
- The customer's own GitHub Actions CI workflow.
- Runs
- One run on each runner: a single example, not an average.
18 stacks we measured ourselves
1.5× median
Median 1.5× faster across 18 measured stacks (15 faster, 2 the same, 1 slower).
- Workload
- First job queued to last job done, full CI run with 10 test shards and a real database.
- Machines
- GitHub's ubuntu-latest (2-core) against our 8 vCPU runner: part of the gain is more cores.
- Runs
- 1 GitHub run and 1–2 compiler.dev runs per stack, measured 2026-10-08.
- Slower
- 1 stack (single-threaded pytest): single-threaded steps gain little from more cores.
Benchmark mode is free, with no card: a copy of your jobs runs on our machines while your results still come from your own runners.
The switch
One line, or one pull request.
Size label
After installing the GitHub App, change the runner line. Sizes 2, 4, 8, 16 vCPU.
- runs-on: ubuntu-latest + runs-on: compiler-8cpu
Routed setup
Merge the setup pull request we open. It routes jobs to us and keeps the runner you use today as the fallback.
Fresh VM per job
Every job starts on a clean machine that is deleted after the job.
Caches ready
Dependency, git and Docker caches are already there, so setup steps take less time.
Spending cap
Set a monthly cap. When it is reached, routed jobs go back to your own runners.
Undo any time
Change the line back or uninstall the App. There is no contract.
Fallback
Keep your existing runners as a safety net.
What happens when something goes wrong depends on how you switched. A real test failure is never silently re-routed.
Routed setup
If compiler.dev can't take a job, or your credit or monthly cap runs out, the job runs on the runner it used before you switched: GitHub-hosted or your own.
Bare size label
runs-on: compiler-8cpu on its own has no automatic fallback: the job waits for a compiler.dev runner.
Real test failures
A job that fails because a test failed stays failed. It is never re-routed to another runner.
Plans
Start on Basic. Move up only if you need to.
| Runner label | vCPU | Per minute |
|---|---|---|
compiler-2cpu | 2 | $0.002 |
compiler-4cpu | 4 | $0.004 |
compiler-8cpu | 8 | $0.008 |
compiler-16cpu | 16 | $0.016 |
Basic
Per-minute prices above, from prepaid credit.
- Turbo and actions cache on the box
- CI insights dashboard
- AI reviewer
- Auto-fix agent
Pro
Runner minutes at the same per-minute prices, plus a monthly plan bought in the app; the price is shown before you pay.
- Bare-metal warm runners
- CI timings insights
- Flake quarantine
- Timing-based sharding
- Visual regression tests
- Merge queue batching
- Toolchain-swap PRs
- Affected-test selection
- Hyper whole-box runners
Turbo
Machines reserved for you. Capacity is limited and set up by hand, so talk to us first.
Stop waiting. Start shipping.
200 free minutes. No charge until they run out; to keep going you add a card and set a monthly spending cap. Benchmark mode is free, with no card: a copy of your jobs runs on our machines while your results still come from your own runners.