There are two ways to connect a cluster, and you can use either or both.
- Managed clusters — AKS, EKS and GKE — are read through the cloud provider's own management API, using the same read-only cloud connection you already set up. Nothing is installed in the cluster and we hold no kubeconfig. (EKS and GKE ride on the AWS and GCP connectors, which are in Beta: verified against emulated APIs, not yet validated on a live customer account.)
- Any distribution — OpenShift, k3s, self-hosted, and managed clusters too — can be connected with a read-only ServiceAccount you create from the manifest on the Connectors page. This in-cluster reader is in Beta: it is tested against a simulated API server and has not yet run against a production cluster.
1. Cluster posture (managed clusters)
- Kubernetes version, whether an upgrade is available, and how close the version is to end of support.
- Cluster and node-pool provisioning state; node counts (for GKE, each pool's configured initial size) and autoscaling bounds.
- Node image or AMI age relative to the control-plane version.
2. Security configuration (managed clusters)
- AKS — Kubernetes RBAC disabled, no Entra integration (or the legacy kind), network plugin.
- EKS — a public API endpoint (and whether it is open to
0.0.0.0/0), control-plane log types switched off, no IAM OIDC provider for service accounts, no KMS envelope encryption for secrets. - GKE — private-cluster setting, node-pool autoscaling and health.
3. Inside the cluster (in-cluster reader, Beta)
The ServiceAccount can only get and list nodes, namespaces, pods, service accounts, resource quotas, deployments, statefulsets, daemonsets, ingresses, network policies, RBAC roles and bindings, cert-manager certificates, and Velero backups and schedules. From those we flag: a version past upstream end of life, nodes not ready or under pressure, kubelet version skew, unavailable workloads, CrashLoopBackOff and image-pull failures and high restart counts, :latest or untagged images, missing memory limits, privileged, host-namespace or hostPath workloads, namespaces with no NetworkPolicy, Ingresses without TLS, cert-manager certificates nearing expiry, cluster-admin bindings to non-system subjects, wildcard roles, a default service account that mounts its token, ResourceQuota usage at 80% and 95%, and Velero schedules with no recent completed backup.
The API server has to be reachable from the public internet over HTTPS, and TLS verification is always on. A private cluster cannot use this reader; if it is AKS, EKS or GKE it is still covered at the management-API level. An outbound-only agent for private clusters is built but not offered yet.
What the Kubernetes connector does NOT do
- Never executes inside your pods. No
kubectl exec, no port-forward, no log streaming. - Never reads Secrets or ConfigMaps. The in-cluster role does not grant them; an Ingress's TLS secret name is read from the Ingress itself.
- Never modifies your cluster. The role has no write verb, and no
watch. - Does not read cluster events.
Clusters are scanned every hour: node readiness, workload availability and restart loops are exactly the kind of state that changes between one nightly pass and the next. An account administrator can ask for a fresh scan of a connection, at most once every 15 minutes per account; it runs with the next hourly scan, and while CloudMonitor is pre-launch nothing scans during the part of each day the service is powered down (see the Terms). The manifest and setup steps are on the Connectors page.