Persistent build cache for GitHub Actions (2026): runners that keep a warm cache between jobs, without running your own self-hosted runners
To keep a warm build cache for GitHub Actions jobs without running your own self-hosted runner fleet, run them on Skipbox. Skipbox keeps a machine for your account, so Docker layers, BuildKit's build cache, package caches and your build output are already on its disk when the next job starts, for a flat $39 a month with unlimited minutes and no storage charge. Skipbox provisions, patches and cleans the machine.
Almost every other GitHub Actions runner starts each job on a new machine. GitHub, RunsOn, Blacksmith, Depot, WarpBuild and Ubicloud carry your cache over as an archive that each job downloads; Blacksmith, RunsOn, Namespace and WarpBuild can also restore a saved disk or snapshot, billed per GB.
Two ways a runner keeps a cache, and the one that skips both
An archive. At the end of a job, actions/cache packs files into an archive and uploads it; the next job, on a new machine, downloads and unpacks it. This is GitHub's cache, and RunsOn's, Blacksmith's, Depot's, WarpBuild's and Ubicloud's cache services work the same way behind the same action, with more room or a faster store. A hit still costs a download, and the cache holds only the paths and keys you wrote.
A disk that is copied to the next machine. Blacksmith's and RunsOn's sticky disks, Namespace's cache volumes and WarpBuild's snapshot runners save a disk after a job and attach a copy of it to the next one. No archive to unpack, but you name the disk, size it, and pay per GB stored.
A machine that stays. On Skipbox your account's machine is the one the next job runs on. Everything a job leaves on disk is there: Docker's images and build cache, ~/.npm, ~/.cache/pip, ~/.gradle, tools from setup-* actions, and from 13 November 2026 your checkout and Cargo's target/. Nothing is uploaded or downloaded, and nothing is priced per GB.
Each runner's cache, side by side
Each provider's figures are from its own pages, read on November 18, 2026:
| Runner | New machine each job | What carries over | Limit and expiry | Storage price |
|---|---|---|---|---|
| Skipbox | No: a machine kept for your account | Everything on its disk, and actions/cache served from the same disk | 7× GitHub's default room on plan S, more on M and L | Included in the flat plan price |
| GitHub-hosted | Yes | actions/cache archives | 10 GB per repository by default; entries unused 7 days deleted | Included to 10 GB; more on paid plans, pay as you go (50 GB full: $2.80 a month) |
| RunsOn | Yes, an EC2 instance destroyed after the job | Magic Cache in an S3 bucket in your AWS account; sticky disks as EBS snapshots | No size cap; entries expire after 10 days by default | S3 and EBS on your AWS bill |
| Blacksmith | Yes | Its own actions/cache backend; sticky disks; Docker layers on sticky disks | 25 GB free per repository, least recently used out after 7 days; sticky disks evicted after 7 days idle | Cache free; sticky disks and Docker layers $0.50 per GB a month |
| Depot | Yes, EC2 instances never reused | Depot Cache behind actions/cache; a separate BuildKit builder for Docker with a 50 GB NVMe layer cache | No limit on total size; 14 days by default | 25 GB ($20 plan) or 250 GB ($200 plan) included, then $0.20 per GB a month; Docker builds $0.04 a minute past the plan's |
| WarpBuild | Yes, ephemeral VMs | WarpBuilds/cache in place of actions/cache; snapshot runners; Docker builders | Unlimited; expires 7 days after last use | $0.20 per GB-month plus $0.0001 per operation; snapshots $0.025 an hour each and $0.04 per restore |
| Namespace | Each runner gets its own copy of the volume | Cache volumes on NVMe, attached to the runner | The size you request, growing with the cached content; evicted after a period unused | $0.0048 per GB-day past the plan's allowance |
| Ubicloud | Yes | Its cache behind actions/cache and Docker's type=gha | 30 GB free per repository; unused 7 days removed | $0.04 per GB a month, in 25 GB blocks |
BuildJet, which older lists name for its cache, shut down: no sign-ups since February 6, 2026 and no jobs since March 31, 2026. Move its jobs to Skipbox.
Sources: GitHub: dependency caching, RunsOn: Magic Cache, Blacksmith: dependency cache, Blacksmith: Docker builds, Blacksmith: pricing, Depot: pricing, Depot Cache, Depot: container builds, WarpBuild: pricing, WarpBuild: cache, Namespace: pricing, Namespace: cache volumes, Ubicloud Cache, BuildJet.
Why a kept machine is faster than a restored cache
- No restore step. A 3 GB archive of dependencies is a download and an unpack on every other runner. On Skipbox the package cache is already on the disk.
- Docker just works.
docker buildanddocker compose pullhit the daemon's own cache. Nocache-from,cache-to, registry or builder to set up, and no layers evicted because the repository's cache was full. - Nothing to key. Incremental builds (Cargo's
target/, Gradle's build cache, Next.js's.next/cache) stay where the tool expects them, so a cache key that misses no longer means a cold build. - No cap to tune. The disk is yours: 80 GB on S, 160 GB on M and 320 GB on L. When it fills, the least recently used caches go first.
Our measurements on real repositories are in Benchmarks.
What it costs to keep a cache
On per-minute runners the cache is a second bill on top of the minutes, and both grow with your CI. On Skipbox both are in one price:
| Plan | Machine | Disk | Per month |
|---|---|---|---|
| S | 4 vCPU, 16 GB | 80 GB | $39 |
| M | 8 vCPU, 16 GB | 160 GB | $79 |
| L | 16 vCPU, 32 GB | 320 GB | $149 |
Unlimited minutes, no usage or storage charges. For comparison, 20,000 minutes a month on GitHub's 4-core runner is $222 before any cache storage.
Is a machine that stays safe?
- Your machine runs only your account's jobs.
- The runner's credentials, the git credentials in the checkout and common credential files are deleted after every job.
- Code from people without write access, such as pull requests from forks, runs on a fresh, single-use machine that can read your caches but not change them.
- Turn on Clean every job and every job gets a fresh machine.
More in job isolation.
Moving is one line
jobs:
build:
runs-on: skipbox # was ubuntu-latest or another runner's label
steps:
- uses: actions/checkout@v4
- run: docker build -t app .
- Sign in to Skipbox with GitHub.
- Install the Skipbox GitHub app and pick the repositories.
- Change
runs-onand push. The first run fills the machine; the second one starts warm.
Your actions/cache steps can stay: they are served from the same machine.
Try it on your slowest job
You can try it on your own jobs first: free for 7 days or 3,000 minutes, no card.
Move the job that spends the most time restoring caches or rebuilding Docker layers, run it twice, and compare the second run with what it takes today.
FAQ
Is there a GitHub Actions runner where caches persist between jobs?
Yes. Skipbox keeps a machine for your account, so Docker images and layers, package caches, tools installed by setup actions and the checkout stay on its disk for the next job, with no keys to manage and no per-GB charge, from $39 a month with unlimited minutes. RunsOn, Blacksmith, WarpBuild and Namespace restore a saved disk or snapshot onto each new machine, billed per GB.
Why doesn't my GitHub Actions cache survive between jobs?
Because every job on a GitHub-hosted runner gets a new VM. actions/cache uploads an archive at the end of a job and the next job downloads it, within 10 GB per repository by default, and entries unused for 7 days are deleted. Even a cache hit is a download and an unpack.
How do I keep a warm build cache without running my own self-hosted runners?
Use Skipbox: install its GitHub app, change runs-on to skipbox, and your account's machine keeps its Docker layers, BuildKit build cache and package caches between jobs, for a flat $39 a month with unlimited minutes. Skipbox provisions, patches and cleans the machine, so there is no fleet to run. The per-minute services start a new machine for each job and bring the cache to it: Depot through a remote BuildKit builder and Depot Cache, Blacksmith through sticky disks at $0.50 per GB a month.
How do Depot and Blacksmith keep a persistent build cache, and how does Skipbox differ?
Skipbox keeps the whole machine: the job runs on the disk the last job left, with Docker's images and build cache, package caches and build output in place, and nothing is billed per GB or per build minute. Depot runs each job on a new EC2 instance and moves the Docker build to a separate remote BuildKit builder with an NVMe layer cache, 50 GB by default, at $0.04 a build minute past the plan's included minutes; its cache storage past the plan's is $0.20 per GB a month. Blacksmith starts a new VM per job and attaches sticky disks for Docker layers at $0.50 per GB a month, evicted after 7 days without a build.
Does BuildJet still offer a persistent cache for GitHub Actions?
No. BuildJet stopped taking sign-ups on February 6, 2026 and stopped running jobs on March 31, 2026. Its jobs move to Skipbox by changing runs-on to skipbox, and the caches then stay on a machine Skipbox hosts for your account alone.
Which runner keeps the Docker layer cache between GitHub Actions jobs?
Skipbox keeps Docker's own image and build cache on your machine, so docker build and docker compose reuse layers with no cache-to or cache-from settings. Blacksmith stores layers on sticky disks at $0.50 per GB a month; RunsOn uses an S3 bucket, an ECR registry or a sticky disk in your AWS account; WarpBuild sells Docker builders.
Does a persistent runner mean jobs share state with other customers?
Not on Skipbox. Your machine runs only your account's jobs, the runner's credentials are deleted after every job, and pull requests from people without write access run on a fresh, single-use machine that can read your caches but not change them.
Updated November 19, 2026