Architecture · Enterprise security
Complete Kubernetes operation without exposing your API server.
The KubeBolt agent opens the outbound connection to the control plane. You get a web terminal, logs, runtime filesystem inspection, port-forwarding and the full kubectl surface — all from your browser. No VPN, no jump host, no public API server.
01 — Access model
Two access models. Only one works on private clusters.
Traditional dashboards have to reach your API server. KubeBolt flips the direction: the agent goes out, nothing comes in. That changes the whole security model.
Traditional model
Lens · Kite · K9s · direct kubectl
- API server must be reachable from the client
- Credentials on every user machine
- Private clusters need VPN or jump host
- Audit lives in the cluster, not centralized
KubeBolt model
Outbound agent — cluster stays private
- API server stays private, never exposed publicly
- Credentials only on the agent (Workload Identity)
- Works on private GKE/EKS, air-gapped, isolated VPCs
- Audit centralized at the KubeBolt control plane
02 — Agent-based capabilities
Everything you'd do with kubectl. From the browser.
The agent is kubectl-complete. Not a dashboard with limited operations: it's the full Kubernetes API surface, proxied through the agent.
01
Interactive terminal on any pod
Exec sessions with full stdin/stdout/stderr. Multi-pod, multi-cluster, straight from the browser. No kubeconfig on the client, no VPN, no access to the API server.
kubectl exec via agent02
Filesystem inspection at runtime
Browse the filesystem of running pods: directories, files generated by the app, configuration loaded at boot, certificates, mounted volumes with real data.
More useful than inspecting the static image: you see what's happening now, not what shipped.
runtime fs · not image03
Multi-pod logs streaming
Real-time streaming with filtering, search and follow. Cross-pod aggregation in replicated deployments. No waiting for the log to reach your observability stack.
live tail via agent04
Port-forward without exposure
Access to internal UIs (Grafana, ArgoCD, private tooling) without LoadBalancers or public Ingress. The tunnel goes through the agent, the services stay private.
tunneling via agent05
Full resource CRUD
Create, update, delete, scale, restart, rollout, all via UI or API. Inline YAML editing with validation. CRDs supported natively.
full kubectl surface06
Multi-cluster by default
One agent per cluster, all visible from the same control plane. No kubeconfig juggling, no mental context-switching. The cluster is just another filter in the UI.
fleet-native03 — Comparison
What a dashboard does, plus what no dashboard has.
Skyhook (the company behind Radar) is a lean 4-person startup with ~$3M raised post-YC W23, competing to build the best modern open-source dashboard. Kite is Lens done well by a single maintainer. KubeBolt is a different category: when the dashboard isn't enough and you need autonomous operation without exposing infrastructure.
| Capability | KubeBolt | Radar | Kite | Lens |
|---|---|---|---|---|
| Outbound agent (private cluster, no VPN) | ✓ | — | — | — |
| Web terminal | via agent | — | direct | direct |
| Filesystem inspection | runtime | image | — | — |
| Multi-pod logs streaming | ✓ | ✓ | ✓ | ✓ |
| Port-forward without public exposure | via agent | local | kube proxy | local |
| Live topology view | roadmap | ✓ | basic | basic |
| Persistent event timeline (>1h TTL) | roadmap | ✓ | — | — |
| Helm releases + ArtifactHub install | roadmap | no ArtifactHub | ✓ | manual |
| GitOps (ArgoCD/Flux) sync state | roadmap | ✓ | — | — |
| AI agentic (not just chat) | Copilot + Autopilot | MCP read-only | chat | — |
| Deterministic runbooks + LLM hierarchy | six-layer | — | — | — |
| Autonomous incident resolution | Autopilot | — | — | — |
| Autonomous FinOps optimization | Lifecycle Mgmt | OpenCost view | — | — |
| MCP server to IDE (Cursor, Claude Code) | ✓ | read-only | — | — |
04 — What KubeBolt actually is
"KubeBolt isn't a dashboard with AI added on. It's an AI operations platform for Kubernetes.
Your cluster self-diagnoses through deterministic runbooks and AI agents built on the Claude Agent SDK. It resolves incidents on its own when it can. It saves you money by powering down non-production environments when nobody's using them. And all of that happens without exposing your API server: the monitoring agent opens the outbound connection from your cluster, the cluster stays private.
The UI is excellent. But the product lives in two pieces: the Core (API), home to the correlation engine and the hierarchy of AI agents; and the monitoring agent inside your cluster that brings the facts to the Core without ever exposing anything."
Leafar Maina · Founder, CLM Cloud Solutions