ClusterMind · AI SRE for self-managed Kubernetes

Built for the Kubernetes clusters cloud tools forgot.

AI SRE for Kubernetes you run yourself.

AI SRE for the clusters nobody else covers.

Deep diagnostics for bare metal, k3s, RKE2 and Talos. ClusterMind understands Longhorn failures, stranded StatefulSets and ArgoCD drift, then turns the fix into a reviewed GitOps PR.

Start using ClusterMind Walk through the incident

Incident on prod-bm-01Criticalworker-03 NotReady

02:14

Scheduled run finds 3 critical issuesnode_health · statefulset_stuck_pods · longhorn_stuck_volumes

02:16

One causepostgres-0 is stranded on dead worker-03, so its Longhorn volume loops on attach.

02:17

Proposed fix, waiting on a personForce-delete db/postgres-0, then open a PR setting nodeDownPodDeletionPolicy.Checked first: a healthy Longhorn replica is on worker-02.

Approved by @dana · 02:19 · scoped to prod-bm-01

Denied by @dana. Nothing changed. The finding stays open, and anyone can ask for a different fix in the thread.

02:20

Pull request opened{{ prLine }}

02:23

Checks re-run{{ verifyLine }}

Nine minutes of one night, slowed down.

The same incident, stage by stage. A worker dies at 2 a.m. with a database on it. This is what ClusterMind did, and where it stopped to ask.

checkresult

worker-03 died with db/postgres-0 on it. StatefulSet pods don't auto-reschedule, so its Longhorn volume loops on attach. A healthy replica exists on worker-02.

1Force-delete db/postgres-0 so it reschedules on worker-02

2Open a PR setting nodeDownPodDeletionPolicy

Approval requestNeeds your decision#platform-incidents

Force-delete db/postgres-0 and open PR to infra-gitops

Checked first: a healthy Longhorn replica is on worker-02.

Approved by @dana02:19 · scoped to prod-bm-01

Denied by @dana. Nothing changed. The finding stays open in the thread.

Scoped to one cluster. Deny, and nothing changes.

acme/infra-gitops · #412 · {{ prState }}longhorn: delete stranded StatefulSet pods on node-down

defaultSettings:

- nodeDownPodDeletionPolicy: do-nothing

+ nodeDownPodDeletionPolicy: delete-both-statefulset-and-deployment-pod

statefulset_stuck_pods{{ vLabel }}

longhorn_stuck_volumes{{ vLabel }}

Built for the stack you actually run.

Distributions

bare metalk3sRKE2Taloskubeadm

Around them

LonghornArgoCDPrometheusAlertmanagerSlack

Managed-cloud SRE tools know EKS. ClusterMind knows what happens when Longhorn won't attach after a node dies.

01 Longhorn attach-deadlock loops critical
longhorn_stuck_volumes
longhorn_instance_managers

02 StatefulSet pods stranded on dead nodes critical
statefulset_stuck_pods

03 Nodes left cordoned after an upgrade critical
node_health
node_conditions

04 ArgoCD drift after a hand fix in production medium
argocd_sync_status

Built from real production incidents, not generic troubleshooting prompts.

Each check carries the incident that produced it and the safe next step. 21 checks, 9 of them for self-managed stacks. Deep diagnostics run every 6 hours by default, or on request in any thread.

- name: statefulset_stuck_pods
  description: StatefulSet pods stuck on failed nodes
  severity: critical
  context: |
    Incident ref: PostgreSQL stuck on a dead node for hours
    StatefulSet pods do NOT auto-reschedule.
    Check if data loss is acceptable before force-deleting.

Every fix is reviewable before it ships.

ClusterMind never holds cluster-admin. Two agents with different permissions, and a person between them. No autonomous destructive actions.

On its ownread-only agent

Run scheduled and on-demand checks

Correlate nodes, volumes, pods and sync state

Explain the cause and propose a specific change

Only after a person approvesprivileged agent

Execute the approved change, scoped to one cluster

Open a pull request in your GitOps repo

Re-run the checks and post the result

Know where your cluster data goes.

Raw cluster data goes only to Anthropic, under your key. ClusterMind stores incident records and the agent's write-ups. Those can quote log lines. Connected Slack and GitHub integrations also receive information. The control plane holds your workspace Anthropic key when one is set; RBAC determines what the agent can read, with no redaction step.

Where each kind of data goes
What	Your environment	Anthropic API, your key	ClusterMind
Logs, specs, events, secrets	Read here	Yes, as agent output	May appear in stored evidence and write-ups
Prompts and inference	Sent from here	Yes, billed to you	Never
Incidents, write-ups, approvals, evidence, usage	Created here	No	Yes

Your environment: {{ pr.env }}
Anthropic API, your key: {{ pr.api }}
ClusterMind: {{ pr.cm }}

Keep Prometheus. Keep Alertmanager.

Alerting tells you something is wrong. ClusterMind works out why and proposes the fix. It doesn't try to page you faster.

Could you build this with Claude + kubectl? Yes.

Then you'd also own all of this:

SafetyRBAC boundariesApproval flowMulti-cluster safetyVerification

PlumbingCurated checksGitOps remediationSlack workflowUsage and cost tracking

Built for teams who run their own Kubernetes.

For

Bare metal, k3s, RKE2, Talos and kubeadm clusters

Longhorn for storage, ArgoCD for GitOps

1–10 person platform teams: on-prem, edge, data center, sovereign, homelab-to-production

Teams already running Prometheus and Alertmanager

Not for

Managed-cloud fleets on EKS, GKE or AKS. Cloud-native agents already serve them.

Sub-minute autonomous detection

Remediation with no human approval

Run Kubernetes your way. Keep control of every fix.

Start using ClusterMind Read-only until someone approves. Review how cluster data is processed in our Privacy Policy.