Kubeadapt MCP
Find idle capacity and node savings
Ask an AI tool connected to Kubeadapt MCP how much you pay for unused capacity, which nodes sit empty, and what a smaller or spot node layout would save.
Last updated
Ask your AI tool how much of the bill pays for capacity nobody uses and it splits the cost into workload usage, idle capacity and shared cost, then finds the emptiest nodes, Kubeadapt's recommended node layout and the volumes nothing claims. It can also estimate what moving capacity to spot or ARM, or removing nodes, would save.
This example needs a connection with the Savings data kind. The idle-cost split also needs Costs and carbon.
Ask
plaintext
How much of last month's bill was idle capacity that no workload requested?plaintext
Which nodes on the prod cluster are the least utilized, and what does each cost per month?plaintext
What node layout does Kubeadapt recommend for the prod cluster, and what would it save?plaintext
What would we save by moving half of our on-demand node cost to spot? And by moving 30% of x86 capacity to ARM?plaintext
Which volumes aren't attached to anything but are still billed?plaintext
How much CPU headroom does the prod cluster have at its current growth?What happens
| Tool | What it returns here |
|---|---|
query_costs | cost_split: the fully loaded bill as workload usage, idle capacity nobody requested, and shared and unallocated cost. |
get_capacity_overview | CPU and memory capacity, requested and used, unrequested capacity, CPU growth and days of headroom, node groups, the cluster's monthly cost now and after Kubeadapt's node-group rightsizing, and the most expensive namespaces. |
list_nodes | Nodes or node groups with instance type, spot or on-demand, zone, allocatable, requested and used CPU and memory, pods and monthly cost. Sorted least utilized first on request. |
get_node_optimization | The recommended node layout and its saving, or a what-if with spot_percent (0 to 100), node_scaling_pct (−50 to 100) or arm_migration_pct (0 to 100). |
list_volumes | Persistent volumes, most expensive first, with size, storage class, status, claim and monthly cost; unattached volumes counted separately. |
list_recommendations | Orphaned-volume and spot recommendations with their monthly saving. |
What the answer contains
- Idle cost.
workload_usage_usd,idle_capacity_usd,shared_unallocated_usdandfully_loaded_total_usdfor the window. - Utilization. For each node, allocatable, requested and used CPU cores and memory GiB with utilization percentages, and monthly cost: its current hourly price × 730. Nodes with no usage sample in the last hour have no used figures.
- Headroom. Unrequested CPU and memory, CPU growth per 30 days, days of CPU headroom, and projected CPU and memory use in 30 days.
- Node layout. Current and recommended node groups, instance types and counts, the nodes' monthly cost now and after, and the saving.
- What-ifs. The nodes' monthly cost after the change and the saving of each lever, signed: a negative saving means the scenario costs more. Spot assumes 35% off on-demand prices for every node moved; ARM assumes 20% off for every x86 node moved. With both set, the discounts are added rather than compounded, so the saving is overstated.
- Volumes. Each volume's size, storage class, status, claim namespace and name, and monthly cost; the summary counts unattached volumes (bound to no claim but still billed) and what they cost.
Follow-ups
plaintext
Show CPU and memory use for the emptiest node over the last week.plaintext
Which workloads in the most expensive namespace have rightsizing recommendations?plaintext
Which spot recommendations are lowest risk, and what do they save together?plaintext
Open the capacity planning view for this cluster in Kubeadapt.The Find idle spend ready-made prompt runs this whole review in one step. See Ready-made prompts.
Limits
- Read-only. Kubeadapt MCP doesn't resize node groups, move nodes to spot or delete volumes. Make those changes through your own infrastructure tooling.
- What-ifs are estimates. Node scaling changes cost in proportion to the node count and doesn't check that the pods would still fit. ARM needs container images that run on ARM.
- On-premises clusters get no recommended node layout, and spot and ARM what-ifs don't apply to them. Node scaling still estimates a cost, which is saved only if those machines stop being paid for.
- The recommended layout needs a recent analysis. If none is recent, or no cheaper layout fits the workloads, the answer says so instead of showing one.
- Volumes link to claims, not workloads. Kubeadapt shows which claim a volume belongs to, not which workload mounts it.
- Up to 100 nodes, node groups or volumes per answer, 25 by default.