KRR (Kubernetes Resource Recommender) is a command-line tool from Robusta that reads historical CPU and memory usage from Prometheus and recommends new requests and limits for every workload in a cluster. It is agentless: nothing is installed in the cluster, the CLI just queries Prometheus and the Kubernetes API directly.
By default it uses the “simple” strategy — CPU request set to the 95th percentile of observed usage, memory request set to the maximum observed usage plus a 15% buffer, and CPU limits unset to avoid throttling latency-sensitive workloads. The strategy is pluggable, so you can ship your own scoring function if you disagree with those heuristics. Output includes a terminal table, JSON, CSV, or HTML, showing current vs recommended values and the cost savings if you applied them. KRR does not write back to the cluster by default; you are meant to review the diff and roll it into your GitOps repo.
It is an alternative to the built-in Vertical Pod Autoscaler (VPA) recommender, which requires installing the VPA controller and is awkward to run in report-only mode. If you already run Prometheus or a compatible store (Thanos, Mimir, Victoria Metrics, Coralogix, Grafana Cloud), KRR is the quickest way to get a ranked list of over- and under-provisioned workloads.