“We were essentially able to reduce the cost of that cluster by about 75%. On AWS, DevZero demonstrated they could achieve significantly higher savings than we initially thought possible.”
Mihir Nair
Head of Architecture, Databahn
Free CLI. Nothing leaves your machine
Visualize the waste in your cluster
- Gets cluster metadata from
kubectlctx - Does math and scheduling simulation locally
- Prints the gap between
requestsand what's used
Not the one with kubectl access?
~ — npx devzero@latest analyze-cluster
~ $
Utilization
What you use,next to what you requested
With metrics-server you get live CPU and memory next to what workloads requested. Without it, the report compares requests to allocatable and says so. That gap is usually the money.
11-node cluster · AWSvia metrics-server
CPU8% used
requested 80%
3.8 of 47.1 cores used · 37.8 requested
Memory31% used
requested 75%
40 of 128 GiB used · 96 requested
The gap between the green bar and the requested marker is capacity you pay for but don't use.
Today11 nodes
repacked, honoring taints, affinity & PodDisruptionBudgets
After8 nodes (−3)
Consolidation
A bin-packing simulation,not a guess
The scan re-packs your actual pods onto fewer nodes and names the ones that could be drained. Taints, tolerations, node affinity, and PodDisruptionBudgets are honored. What it can't model, like anti-affinity and topology spread, gets flagged instead of assumed away.
Savings estimate
A dollar figureyou can defend
Public on-demand pricing, split into rightsizing and spot-conversion headroom, capped so it can never exceed your current bill. It reads conservative on purpose. It has to survive a budget meeting.
Current est. spend$762–$1,197/mo
Rightsizing headroom$604–$955/mo
Spot-conversion headroom$38–$86/mo
Est. savings with DevZero$642–$1,041/mo
Public on-demand list pricing, based on actual usage scaled to a ~70% utilization ceiling. Hard-capped at your current spend.
The trust model is deliberately boring
It exists for the moment before you'd give a vendor anything. So it takes nothing.
~/.kube/config
foundcurrent-context: prod-eks
DevZero account
not neededno sign-up flow exists
API token
not neededno --token flag to forget
Payment details
not neededit's a free scan
Your kubeconfig is the only credential
No token, nothing to sign up for. It uses the access your kubectl already has: read-only, same trust model as kubectl top.
network activity during a scan
kube-apiserverGET nodes, pods, pdbs112 reqs
metrics.k8s.ioGET node & pod usage22 reqs
api.devzero.io—0 reqs
telemetry—0 reqs
Zero bytes leave your machine
Collection, simulation, and pricing all run locally. No telemetry, no phone-home, not even a version check.
verbs it uses
getlistwatch
verbs it never asks for
patchupdateevictdraindelete
Consolidation is simulated in memory. Nothing is drained for real.
It can't change anything
It only reads. Draining, rightsizing, and spot conversion are estimated, never executed. Your cluster is untouched.
Not the one with kubectl access?
Ask your platform engineer.
One command, a minute of their time. Copy the ask and drop it in Slack.
youto #platform-eng
Hey, can you run npx devzero@latest analyze-cluster against prod and send me the report? It's read-only, runs locally, and takes about a minute (found it on devzero.io/analyze-cluster if you want to vet it first). Curious what our utilization actually is.
What our customers say
Frequently asked questions
Things platform engineers asked before pointing it at a real cluster.
Does it send anything to DevZero?+−
No. It runs against your current kube-context, and the collection, simulation, and pricing all happen on your machine. There's no account, no token, and no network call to DevZero anywhere in the code path.
Do I need metrics-server installed?+−
It helps, but it isn't required. With metrics-server you get live CPU and memory usage. Without it, the report falls back to comparing requests against allocatable capacity and labels the output accordingly, so you always know which mode produced the numbers.
Which clusters and clouds does it work with?+−
Anything your kubectl can reach. Spot vs. on-demand detection reads the node labels that Karpenter, EKS, GKE, AKS, and OCI already set. GPU capacity comes from node resources, so GPU nodes show up too.
How seriously should I take the savings number?+−
As a starting point, not a quote. Prices come from public on-demand list pricing, the basis is printed under the number, and the total is clamped so it can never exceed your current estimated spend. It's meant to tell you whether a rightsizing conversation is worth an hour of your time.
Is it safe to run against a production cluster?+−
Yes. It's read-only and never drains, evicts, or modifies anything. The consolidation figures come from a simulation that honors taints, tolerations, required node affinity, and PodDisruptionBudgets. The load on the cluster is about the same as a kubectl get across your nodes and pods.
Can I get machine-readable output?+−
Yes. Run it with -o json for structured output, or -o plain to skip the panels. The --target-utilization flag (default 70%) sets the sustainable-utilization ceiling the rightsizing math aims for.
The report takes about a minute.The waste has been there for months.
If the numbers are real money, DevZero can go collect them. Continuously, with zero restarts.