CRITICAL payments-api
CrashLoopBackOff · 7 restarts
Kobi Autopilot · autonomous remediation for Kubernetes
Your cluster breaks at 3am. You sleep through it.
Kobi Autopilot finds the cause, applies a fix from a closed catalog of verified actions and reverts it if the check fails. You decide how far it goes: start in suggest-only mode and raise the level as you gain trust.
In our own production: real incidents resolved with no human intervention.
A pod enters CrashLoopBackOff at 03:14; Kobi diagnoses the liveness probe, applies a catalog fix and verifies it in 87 seconds. The pager goes from critical to resolved. One time in three it is another incident, an OOMKilled on checkout-worker: Kobi raises the memory limit, verification does not pass because it is a leak, so it reverts the change and escalates to you.
AI agent · 24/7
01 · The problem
The alert tells you what broke.
It doesn't tell you what broke it.
From the alert to the commit that caused it, hopping across five tools. The same incident, side by side, with Kobi.
The same incident, side by side. You: from 03:14 to 03:34, across five tools, until you find the commit that caused it. Kobi: 87 seconds to diagnose it, fix it with a catalog action, verify it and write the postmortem.
You · 5 tools · 10 tabs 20 min to learn what changed. And it still isn't fixed.
Kobi · closed catalog · 0 tabs 87 s to fix it, verify it and have the postmortem written.
At 03:34 you know what broke it. Kobi had fixed it at 03:15.
Reconstruction of a typical incident · illustrative data · real scale: 20 min vs 87 s
Time
Twenty minutes to learn what changed. The alert doesn't say; five tools each know half of it.
Access
Access, scattered. An up-to-date kubeconfig, a VPN or a bastion per cluster, and again for every new person.
Context
Ten tabs, zero context. None of them links the event to the deploy and to the permission that explains it.
02 · How it resolves
The model doesn't decide
what can touch production.
Deterministic rules identify the covered cases and when an investigation is needed. When Kobi proposes an action, the executor only accepts operations from a closed catalog and honors the cluster's RBAC. The only thing that changes is who approves.
▶ Try it: pick a level, break something and see how far Kobi goes.
1Pick the control level
2Break something
three get resolved · one reverts · one stays out
Kobi Copilot · approve and run
Kobi diagnoses, shows you the exact command and runs it on your click. Destructive actions always carry your signature.
Who actsKobi + you
Who approvesYou press Run
SpeedSeconds to the fix
Same engine: detect, diagnose, remediate. Only the control model changes.
Kobi payments-api · CrashLoopBackOff Approve and run
03:14:07 Alert CRITICAL · payments-api · CrashLoopBackOff · 7 restarts
03:14:19 Diagnosis commit #4127 raised startup from 8 to 41 s and the liveness probe kills the pod before it finishes · confidence 0.91
03:14:48 Proposal patch_probe payments-api · 90 s startupProbe
patch_probe payments-api · 90 s startupProbe
$kubectl patch deployment payments-api -n payments -p '{"spec":{"template":{"spec":{"containers":[{"name":"api","startupProbe":{"httpGet":{"path":"/healthz","port":8080},"periodSeconds":10,"failureThreshold":9}}]}}}}'
closed catalogreversiblenon-destructive
Approve and run
Kobi shows you the exact command and runs it on your click, with your RBAC and an audit log.
signed · you waited 00:04
03:15:06 Execution startupProbe applied · 3/3 pods Ready in 34 s · run on your click
03:15:38 Verification pods Ready · no restarts in 60 s · p99 back to 180 ms
03:15:40 Postmortem postmortem written · links commit #4127 and who merged it
Resolved and verified in 91 s. With your signature where it was needed.
Illustrative script using actions from the real catalog · timings follow the 03:14 incident.
Closed catalog · cluster RBAC · confidence threshold per rule · automatic rollback · everything is logged
03 · What connects
Outbound, never inbound.
What you install and what you don't.
The five answers a platform team asks for before installing anything in production.
An agent that dials out, never in
Outbound gRPC channel from inside the cluster. Your API server is never exposed: no VPN, no bastion, no open ports.
ServiceAccount and three RBAC levels
No kubeconfig on the operator's laptop. Start with metrics or read-only and move up to operator whenever you want.
Suggest only, by default
No action runs without your policy. Closed catalog; destructive actions require human approval by platform policy.
Already running Prometheus? Keep it
If you have it, it's joined to resources, events and changes. If not, the agent uses kubelet metrics.
Which data leaves and where it's processed
Only what's needed to investigate leaves the cluster. What is sent, how long it's kept and where, in the trust model.
YOUR CLUSTER · GKE, EKS, AKS, ON-PREM
KubeBolt agent
ServiceAccount · no kubeconfig
RBAC: metrics ▸ read ▸ operator
starts read-only
acts per RBAC
API server
not exposed · no VPN
reads metrics
Prometheus
yours, if you have one
keep operating, unchanged
workloads · events · logs · changes
outbound gRPC · TLS nothing comes in
YOUR NETWORK · FIREWALL
KubeBolt Cloud
- Kobi · Copilot and Autopilot
- dashboard · action log
- closed action catalog
or self-hosted on your infra
04 · Kobi, one agent, two modes
Kobi resolves.
With you in Copilot. Alone in Autopilot.
The same agent, the same closed catalog, the same log of every action. The only thing that changes is who approves.
Kobi Copilot with you · available today
Why has payments-api been restarting since 03:03?
querying your cluster get_workload_pods payments-api get_pod_logs payments-api get_workload_history payments-api
CrashLoopBackOff since the 03:02 rollout. The ConfigMap payments-config changed DB_URL and the pod can't connect. 7 restarts, no impact on other services.
root cause the payments-config ConfigMap, in the 03:02 rollout · confidence 0.94
Proposed action · catalog
rollback_deployment payments-api → rev 41
Copilot: you ask, Kobi answers. Managed in Cloud, your own key if you self-host, and an MCP server.
Kobi Autopilot Open beta
Orchestrator dedup · per-org semaphore · circuit breaker
- 03:14:07 L1 Detector known pattern: CrashLoopBackOff · 7 restarts spotted by a rule, recommendation included
- 03:14:09 L2 Router signal, not noise · escalates to investigation
- 03:14:19 L3 Investigator liveness probe kills the pod before it starts · confidence 0.91
- 03:14:48 L4 Planner patch_probe from the catalog · reversible
- 03:15:02 L5 Executor applied under guardrails · not destructive, no approval
- 03:15:34 L6 Postmortem pods Ready · report written
deterministic AI + deterministic model
Autopilot only handles catalogued incidents.
Autopilot: Kobi acts alone, you approve destructive actions. Watch the incident · 1:40
You choose the level. Suggest only · Approve and run · Autonomous. Destructive actions always need your signature.
Closed catalog + RBAC. Verified, reversible actions. No arbitrary kubectl.
Everything is logged. What it did, the outcome and who approved it. If verification fails, it undoes its own change.
In our own production: real incidents resolved with no human intervention.
05 · The platform
Health, security and spend.
Nothing else to install.
The same agent that already connects the cluster opens four views. Today, on Free.
Cluster: Overview · Capacity · Cost · Reliability
Home. What happened while you were away, what needs your attention today and the spend, across every cluster.
Fleet. Each cluster's health, nodes, pods and spend, at a glance.
Security. Vulnerabilities, RBAC and CIS from your own scanners.
Cluster · Cost. OpenCost, idle and recommended rightsizing, in beta.
06 · Open Source and Cloud
Open Source is free forever.
Cloud charges for infrastructure, not people.
KubeBolt ships in two editions. Open Source is Apache 2.0, unlimited, on your own infra; KubeBolt Cloud adds managed Kobi Autopilot and team access with OAuth, with a free tier available now.
Open Source · Apache 2.0
No cost · you run the infra
The same agent, under Apache 2.0 and unlimited, running on your own infrastructure. Includes Kobi Copilot with your own key (BYOK).
View on GitHubSaaS · Cloud
$0 $49€45 /mo + usage
The same agent, with managed Kobi Autopilot and team access via OAuth. Enterprise Self-Hosted is this same SaaS in your infrastructure.
Start free →Included in the plan
Free Team
Clusters 2 unlimited
Nodes / pods 10 / 150 usage-based
Users 3 (hard cap) 15 (soft)
Retention 15 days 30 days
AI credits / mo 100 1,500
What is charged today
Business, Enterprise and the calculator During launch, overage on nodes, pods and series is metered but not billed; AI credits pause when they run out.
Connect your cluster
in minutes.
Create the account, install the agent, and on the first alert you'll see Kobi working with you. Start in suggest-only.
Free · 2 clusters · no time limit
Or install the open-source agent
$ helm install kubebolt-agent \
oci://ghcr.io/clm-cloud-solutions/kubebolt/helm/kubebolt-agent \
--namespace kubebolt-system --create-namespace \
--set backendUrl=agent.kubebolt.io:443 --set tls.enabled=true \
--set auth.mode=ingest-token \
--set auth.ingestToken.existingSecret=<secret-with-your-token> \
--set cluster.name=<cluster-name>The in-app wizard generates this command with your token and cluster name. docs/connect-clusters
Apache 2.0 · ★ on GitHub · the agent dials out, never in.
What's next
One note per release, no noise.