Skipbox offers flat-price GitHub Actions runners with a warm cache

A CI bill grows with every push. GitHub Actions bills per minute and each run starts on a fresh machine. On 11 November, a Tokyo developer launched a runner service built on the opposite choice. It keeps the machine between jobs.

Skipbox is a hosted GitHub Actions runner service. Its maker, Tomoya Fujita at Fluxenic Inc., launched it on Product Hunt on 11 November. The launch post had 72 upvotes when checked. A team switches to it by installing a GitHub App. They then change runs-on: ubuntu-latest to runs-on: skipbox in one line of the workflow.

What is different

Skipbox runs each job on a Linux VM that persists between jobs. Between runs, it keeps the Docker layer cache, package caches and installed tools. After each job it wipes the credentials, the workspace and any containers. A later job cannot read secrets from the one before. Fork pull requests always run on a fresh machine. There is a setting that gives any job a fresh machine too.

Skipbox says the cache space is "at least 7x" what a GitHub-hosted runner gives you, at 80 GB of disk. The CPU is 2x that of ubuntu-latest, at 4 vCPU. The company says it is "1/12 the cost of GitHub at full use" and "1/8 Blacksmith's". These are the company's own figures.

What it costs

The S plan is $39 a month. It gives 4 vCPU, 8 GB of RAM, 80 GB of disk and one concurrent job. The M plan is $79 a month for 8 vCPU, 16 GB and two concurrent jobs. The L plan is $149 a month for 16 vCPU, 32 GB and four concurrent jobs. All three plans carry unlimited minutes. Yearly billing is ten months' cost, and the free trial is seven days with no card. Skipbox is private repositories only, so an open-source project cannot use it on a public repo.

The pitch is a flat bill. A team pays per minute on GitHub Actions today. Its build spends most of its time pulling the same Docker layers each run. That team gains the most. The gain shrinks for a team whose build is CPU-bound rather than cache-bound. Skipbox offers a money-back guarantee on jobs that run slower on its runner than on GitHub's. The test uses the team's own workload.

What is not said

Skipbox gives no public benchmark on a named open workload. The 1/12 and 1/8 comparisons are not tied to a sample build. A reader cannot check them against a repository they know. The trial lets a team measure the gain on its own build before any card is entered.

Source

Skipbox on Product Hunt. Pricing and feature claims from skipbox.dev.