reap is a static linter for your Kubernetes ML manifests. It flags idle GPUs, missing limits, and runaway jobs right in your pull requests — with the dollar cost attached. No agents. No cluster access. Runs in CI.
The expensive mistakes ship as YAML: a job with no deadline, a GPU request with no limit, an inference pod with no readiness probe. By the time a dashboard shows the waste, you've already paid for it. reap moves the check left — to the pull request, where the config is still cheap to fix.
Jobs that never terminate, notebooks left running, pods sized for peak and idle all night. reap flags the patterns that turn into dollars-per-hour bleed.
Missing shared-memory config that crashes PyTorch/NCCL, absent probes, no gang scheduling for distributed training. Caught before deploy, not after the page.
Baseline the issues you can't fix yet, gate CI on new ones, and enforce one standard across every team's repo — without touching a running cluster.
Drop in the GitHub Action or run helm template . | reap -. It reads your manifests locally — no cluster credentials, ever.
reap comments on the pull request with each GPU risk and its estimated monthly cost. New issues fail the check; known ones stay baselined.
The hosted app rolls up findings and dollar exposure across all repos over time — so you can prove the waste you prevented.
Cast AI, ScaleOps and friends want an agent in your cluster, write access to your node pools, a multi-week rollout, and 10–20% of your cloud spend. If you run a handful of GPUs, the fee costs more than the savings. reap is the opposite end of the stack.
| Enterprise autoscalers | reap | |
|---|---|---|
| Cluster agent required | Yes — write access | No — static only |
| Time to first value | Weeks | Minutes |
| Catches waste… | After it's running | Before merge |
| Pricing | 10–20% of cloud spend | Flat, from free |
| Fits < 50 GPUs | ROI doesn't add up | Built for it |
| Open-source core | No | Yes |
Join the early access list. We'll onboard teams in small batches and you'll help shape the rules that matter for your stack.