Dev sandbox and self-hosted CI on a VPS

Docker, Forgejo and throwaway runners on a no-ID KVM: nested virt, custom ISO, live in about a minute, billed in coins. Delete when done.
Spin up a VPS, install Docker and Forgejo from their official docs, point a runner at it, and destroy the box when the branch merges. Provisioning takes about a minute after the coin invoice confirms, and a deleted VM leaves nothing to keep.
What people actually run
The standard lab is Docker docs plus Forgejo — or GitHub-compatible runners against any forge — with the sandbox rebuilt from a compose file whenever it drifts. Nested virtualization is enabled, so kind, k3s or a test VM under the CI job all work. Billing in USDT keeps CI spend boring and dollar-shaped; a token-only signup keeps an inbox off the throwaway account entirely.
VPS from $4.80* — Core on the annual term (list $6/mo), one invoice for 12 months, no auto-charges. Dedicated with IPMI from $39.20/mo. Crypto invoice, no KYC.
Launch nowWhich forge and which runner
Three stacks cover self-hosted CI in 2026; all three run on the same KVM image with no special hosting hooks:
| Stack | Shape | Pick it when |
|---|---|---|
| Forgejo + its built-in Actions runner | Forge and CI in one binary pair, GitHub-compatible syntax | You want one box to run the whole lab |
| Gitea + Woodpecker CI | Lean forge plus a separate pipeline server | Pipelines outnumber repos, or runners scale on other boxes |
| GitLab CE self-managed | The heavyweight: registry, pages, CI config in-repo | You need the full feature set and have the RAM for it — Prime is the floor |
Stack line-up, checked 2026-10-25.
Runner mechanics are identical behind any of the three: a token pairs the runner to the forge, jobs land in containers, and the cache volume is the only state worth backing up.
Playbook
- Core for a toy. Prime / Prime for runners. Dedicated if steal shows up.
- Install from Docker docs and Forgejo.
- Amsterdam if the team is EU. Delete the VM when finished.
Job to box: the sizing table
Match the job class to the tier instead of guessing; prices are catalog list and annual-eff, checked 2026-10-25:
| Job | Tier | List | Annual-eff | Why |
|---|---|---|---|---|
| Lint, unit tests, one queue | Core | $6 | $4.80 | 2 vCPU / 4 GB absorbs a single job slot |
| Parallel test shards, 2-3 runners | Prime | $12 | $9.60 | 4 vCPU / 8 GB; the practical CI floor |
| Merge train of a small team | Titan | $24 | $19.20 | 8 vCPU / 32 GB absorbs container builds plus cache |
| Heavy builds, no steal tolerated | Forge (metal) | $49 | $39.20 | Whole 9950X cores; steal leaves the conversation |
Artifacts and caches belong off-box — a second VPS keeps them from fattening every new sandbox. Custom ISOs mount fine, so exotic build environments are a template problem, not a ticket.
VPS minutes versus CI minutes, costed
A flat box changes the math of pipeline spend: Prime at $9.60 annual-eff runs 720 hours a month — a busy runner's worth — for less than many metered CI products charge for 2,000 build minutes. Throwaway sandboxes sharpen it further: a VM lives for the length of a review, bills nothing while deleted, and the compose file rebuilds it in about a minute. The honest break-even: spiky, low-volume pipelines often stay cheaper on metered CI; steady daily volume past a few hours a day crosses over to the flat box, and once nested virtualization or custom kernels enter the job, metered CI stops being an option at all.
Where sandboxes earn their keep
- Preview stands per branch — a VM per review, compose-file stack, deleted on merge; the env never outlives the diff.
- Nightly test matrices — k3s, kind and nested-VM tests that metered CI prices like a hobby tax.
- Integration environments with real dependencies — Postgres, Redis and the awkward sidecar that SaaS CI sandboxes refuse to run.
- Experiment boxes — a new database or queue tried against production-shaped data, then deleted with its bill.
- Self-hosted runner pools — Prime-sized boxes joined to the forge as throwaway executors with warm caches.
Limits
Two rules bound the sandbox: do not mine on shared KVM as a business, and do not scan the internet from a box that traces to your own account. CI belongs to build pipelines; anything else reads differently in an abuse queue. And three jobs belong elsewhere entirely: builds that need compliance-certified infrastructure should sit with a vendor that sells certifications, not a no-KYC rack; GPU-sized compile farms want metal, not a VPS tier; and secrets belong in your forge's secret store, not baked into images that a deleted VM leaves behind.
Can I mount a custom ISO?
Yes. Custom ISO mounts are standard on the KVM line, which covers exotic build agents and recovery boots without a support ticket.
KYC for CI boxes?
No. A token or an email plus a 12-character password opens the account; the crypto invoice finishes it. No documents, no card, no phone.
How fast is a throwaway VM ready?
Roughly a minute after the invoice confirms — fast enough to model ephemeral runners that exist for one pipeline and are deleted with it.
What happens to the disk after deletion?
The VM and its storage leave the account with it; the host does not retain the disk. Anything worth keeping should be exported off-box or to your own offsite first.
Where do build caches live so they survive the sandbox?
Off the sandbox — point the runner's cache at a second small VPS or an S3-compatible bucket, and every fresh sandbox mounts the same warm cache instead of re-downloading the world.
How should CI secrets be handled on a rented box?
In the forge's secret store, injected into the job at runtime and never baked into images or compose files; a deleted VM leaves nothing, and an encrypted runner token leaves nothing worth stealing.
Primary sources
Ready to launch?
Build the box — VPS or bare metal — create the password, pay the invoice that follows.