StarSling launches GitHub Actions runners with agents that open CI optimization PRs
StarSling pitches itself as AI-native CI for GitHub Actions. The product has two parts: drop-in Ubuntu runners that replace ubuntu-latest and ubuntu-24.04, and AI agents that analyze a customer's workflows, jobs, run logs and machine telemetry, then open pull requests to improve the pipeline. The PRs target caching, parallelization, dependency installs, test sharding, workflow structure, cache fixes, Docker health checks and job ordering. The customer reviews and merges them. The company says that together the runners and agent PRs made GitHub Actions up to 10x faster in its public benchmarks, and it claims 3M+ CI jobs run. The pitch is aimed at repos where coding agents such as Claude Code, Cursor and Codex have pushed pull request volume up several-fold, which raises CI cost and queue times.
Switching takes one line. A workflow changes runs-on: ubuntu-latest to runs-on: starsling-ubuntu-26.04 (or ubuntu-24.04 to starsling-ubuntu-24.04), and StarSling says no other config changes are required. It is not a full replacement for GitHub Actions: you keep your GitHub Actions workflows and only change the runner label, for supported Ubuntu/Linux jobs. Runners are x64 Linux, from 2 to 64 vCPU, on 5th Gen AMD EPYC hardware. For ARM, or for Docker image layer caching, the page tells readers to email founders@starsling.dev.
The page shows three customer case studies, all with figures supplied by StarSling. Mastra's test suite went from 29m 56s to 5m 06s at wall-clock p95, headlined as 6x faster. Better Auth's E2E time went from 2m 22s to 1m 04s, headlined as 2x faster. Partcl's queue time went from 9.5 minutes to 35 seconds, headlined as 16x shorter. Each has a short testimonial from an unnamed company representative. Partcl's says that within a day of migrating, the agents opened a PR that made its Rust CI tests 2x faster.
The latest feature is Adaptive Compute. A Sandbox Router sends each job to the fastest available sandbox, and a feature called Afterburner adds up to 25% more CPU and up to 50% more RAM to a running step at no extra cost. StarSling claims up to 2.32x faster than GitHub-hosted runners. Other products on the page: StarSling Disks, persistent directories mounted into runners so dependency stores and build output such as pnpm's store or Rust's target/ are ready when the next job starts (on Mastra's 4.5 GiB pnpm store, dependencies were ready in about 25 seconds, 4.5x faster than GitHub's actions/cache, at $0.50 per GB-month); GPU runners with a dedicated NVIDIA card per job; and Review Runners, code review agents that run in the customer's own CI using the customer's model key and post a review comment on each pull request.
Pricing: $0.008 per minute for a 4 vCPU, 16 GB RAM Linux runner, against GitHub Actions' $0.012 per minute, which StarSling calls 33% cheaper. New paid subscriptions are billed by the second with no one-minute minimum, queue time is never billed, and new accounts get 2,000 free minutes in their first month. GPU runners are priced at a normalized $0.06422 per minute for RTX PRO 6000, $0.03022 for RTX 5090 and $0.02522 for RTX 4090, with H100 listed as coming soon. Review Runner time is $0.004 per minute on a 4 vCPU, 16 GB runner, and model tokens are billed by the customer's own provider. Optimization PRs are on paid plans and off by default for new accounts.
On security, StarSling says each job runs in a single-use, hardware-isolated microVM destroyed after the run, the same isolation model GitHub-hosted runners use, and that secrets pass directly from GitHub to the job and are never stored by StarSling. Review Runners load skills and instructions from the trusted base branch, so a pull request cannot rewrite its own reviewer.
Key facts
- StarSling offers drop-in GitHub Actions Ubuntu runners plus AI agents that open pull requests to optimize caching, parallelization, test sharding and workflow structure; the customer reviews and merges.
- Switching means changing runs-on: ubuntu-latest to runs-on: starsling-ubuntu-26.04; StarSling says no other config changes are needed, and that it is not a full GitHub Actions replacement.
- Case studies on the page, all StarSling's figures: Mastra test suite 29m 56s to 5m 06s at wall p95 (6x), Better Auth E2E 2m 22s to 1m 04s (2x), Partcl queue time 9.5 min to 35 s (16x).
- A 4 vCPU, 16 GB Linux runner costs $0.008 per minute versus $0.012 for GitHub Actions; queue time is never billed and new accounts get 2,000 free minutes.
- Optimization PRs are on paid plans only and off by default for new accounts; the 'up to 10x' headline claim combines runners and agent PRs.
Why it matters
The pitch rests on a real pressure point: coding agents such as Claude Code, Cursor and Codex push pull request volume up several-fold, and CI cost and queue times grow with it. StarSling's answer is runners with unlimited concurrency, per-minute pricing with no seat fees, and agents that keep tuning the pipeline so it gets faster as volume grows. The distinctive idea is pairing faster hardware with agents that remove redundant work, rather than selling hardware alone.
Who it affects
Teams that already run GitHub Actions on ubuntu-latest or ubuntu-24.04 and feel slow tests or long queues, especially repos with heavy agent-generated PR traffic. The page's customers are Mastra, Better Auth and Partcl. Teams that need ARM runners or Docker image caching are told to email the founders, and teams wanting a full CI/CD platform swap are not the target, since StarSling only replaces the runner.
How to use it
Sign up with GitHub, install the StarSling GitHub App, and change the runs-on label in your workflow files (ubuntu-latest to starsling-ubuntu-26.04, or ubuntu-24.04 to starsling-ubuntu-24.04). New accounts get 2,000 free minutes in the first month, and the page says no credit card is required. GPU jobs use runs-on: starsling-ubuntu-26.04-gpu with a card chosen in the label. Agent optimization PRs need a paid plan and are off by default for new accounts. StarSling Disks cost $0.50 per GB-month. Volume discounts are available.
How solid is it
Every number comes from StarSling's own page. The three case studies are vendor-published, and only the Mastra figure states its metric (wall p95); the page does not say whether the Better Auth and Partcl figures are medians, means or p95s. The 'up to 10x' benchmark claim comes with no methodology or repository set and combines runners with agent PRs. The 2.32x Adaptive Compute figure is also a maximum. The page gives no independent verification, and the testimonials are from unnamed company representatives.
Risks and caveats
'Up to' figures are best cases, not typical results. The 10x number includes agent-authored PRs, which are not available on the free tier by default. StarSling is a runner alternative only, limited to supported Ubuntu/Linux jobs, with ARM runners not offered on the page and H100 GPUs still listed as coming soon. Agents open PRs that you must review and merge, so the speedups depend on that review work. The page says jobs run in single-use microVMs and secrets are never stored, but these too are the vendor's own statements.
“At Mastra we move so fast that the bottleneck becomes reviews and CI. Time spent compounds, and StarSling helps you realize how much time you're losing.”
— Unnamed Mastra representative, quoted on StarSling's website