Skip to content
Start free

Kubeadapt MCP

Rightsize workloads

Ask an AI tool connected to Kubeadapt MCP which workloads are over-provisioned, get recommended requests and limits, and a YAML patch to review.

Last updated

Ask your AI tool which workloads ask for more than they use and it lists Kubeadapt's rightsizing recommendations by monthly saving, then gives the full plan for any one of them: recommended requests and limits per container and a YAML patch to review. In an editor such as Claude Code, Cursor or VS Code, the AI tool can apply that patch to the manifest in your repository for you to review.

This example needs a connection with the Savings data kind.

Ask

plaintext
Which workloads on the prod cluster are the most over-provisioned, and how much would rightsizing save?
plaintext
Get the rightsizing plan for the api deployment in the shop namespace and give me the YAML patch.
plaintext
Update deploy/checkout.yaml in this repository with Kubeadapt's recommended requests and limits for checkout. Don't apply anything to the cluster.
plaintext
Show redis CPU and memory use over the last two weeks against its requests and limits. When does it peak?
plaintext
Was the payments worker OOM-killed recently, and what memory request and limit should it have?
plaintext
Which spot recommendations carry the lowest risk?

What happens

ToolWhat it returns here
list_recommendationsRecommendations sorted by estimated monthly saving: rightsizing, spot and orphaned volumes. Pending ones by default. Without a cluster or type, it returns each cluster's total saving per type.
get_rightsizing_planOne workload's plan: per container, current and recommended CPU and memory requests and limits, p95 and peak use, OOM kills and restarts, the confidence, the saving, and patch_yaml.
get_usage_historyCPU and memory used against requests and limits over time, with the average, p95, peaks, the busiest hour of the day and the trend.
searchThe workload's id from its name. You can also name the namespace and workload directly.
list_best_practice_findingsConfiguration findings such as missing requests, limits or probes, with the cluster's compliance score.

What the answer contains

  • The saving. monthly_saving_usd, current_monthly_cost_usd, target_monthly_cost_usd and saving_pct. Savings are estimates of the monthly run-rate reduction.
  • Per container. For CPU and memory: current_request, current_limit, recommended_request, recommended_limit, usage_p95, usage_max and efficiency_pct, plus oom_kills and restarts.
  • How sure Kubeadapt is. confidence, observed_hours, pods_observed, the analysis window and when it was last analyzed. Confidence is low when the workload was observed for less than 72 hours. Spot recommendations carry a risk level instead, and orphaned volumes a cleanup-safety rating.
  • The patch. patch_yaml sets container resources with CPU in millicores, such as 500m, and memory in Mi or Gi, such as 1536Mi. Memory is rounded up, never below what the recommendation asks for.
  • Warnings worth reading. Notes say when a HorizontalPodAutoscaler manages the workload (lower requests change the utilization it scales on), when the plan raises some resources because the workload uses more than it asks for, when a container was OOM-killed, and when the recommendation is dismissed or applied rather than pending.

Follow-ups

plaintext
Is the recommended memory safe given redis's peaks over the last 30 days?
plaintext
Do the same for every workload in the shop namespace and list them by saving.
plaintext
Which workloads in the shop namespace have no memory limit?
plaintext
Open this recommendation in Kubeadapt.

The Rightsize workloads ready-made prompt runs the namespace-wide version in one step. See Ready-made prompts.

Limits

  • Proposals only. Kubeadapt MCP never changes a workload or a cluster. Apply the patch through your manifests, Helm values or GitOps, and review it like any other change. When an editor-based AI tool edits a file, that is the AI tool working in your repository, not Kubeadapt.
  • Usage history covers up to 60 days, 7 by default, in at most 60 points.
  • Up to 100 recommendations per answer. Without a cluster or type, totals cover the 10 clusters with the highest run rate; name a cluster for the others.
  • Recommendations come from Kubeadapt's analysis. A workload that hasn't been analyzed yet, or was observed only briefly, has no plan or a low-confidence one.

Related