Meet Kobi, your AI SRE. One agent, two modes: Kobi Copilot fixes it with your click, Kobi Autopilot fixes it hands-free.
The spine — How it resolves
The same resolution engine across the board. The only thing that changes is who pushes the button — you, or the cluster itself.
Same detect → diagnose → remediate engine. Only the control model changes.
Kobi Copilot · Available today
Press ⌘ J and describe what's wrong. Kobi queries your live cluster through 17+ tools, finds the root cause, and proposes the exact fix — you click Execute. It explains every step, and when a fix is beyond what it can run, it hands you the precise console commands.
01 · Resolves
Kobi finds the root cause and proposes the exact command — rollback, scale, patch limits. You approve; it executes with full RBAC and an audit entry. The human stays in the loop.
02 · Teaches
Each diagnosis comes with the reasoning and the commands behind it. Your team levels up on Kubernetes — not just on KubeBolt. Onboarding and incident response in one.
03 · Guides to console
When a fix falls outside what Kobi can execute, it doesn't shrug — it hands you the exact step-by-step to run by hand, so you're never blocked.
04 · Anywhere
Kobi exposes a Model Context Protocol server. Plug it into Cursor or Claude Code and your assistant gets live, read-aware cluster context without leaving the editor.
Multi-model
For interactive, human-approved work, Kobi lets you pick your provider: Anthropic Claude, OpenAI GPT, xAI Grok, or any OpenAI-compatible endpoint (self-hosted Llama / Qwen / Mistral via vLLM, Ollama, LM Studio). Set one as fallback on 429 / 5xx. Prompt caching keeps the bill predictable. (Autopilot, acting unsupervised, runs Claude exclusively.)
Anywhere via MCP
Kobi exposes a Model Context Protocol server. Plug it into your IDE and your AI assistant gets read access to your cluster — without leaving the editor.
Kobi Autopilot
Open beta · Free planAutopilot is Kobi acting on its own: it wakes when something matters — a crash loop spreads, an SLO budget burns, a node degrades. It opens a session, gathers context, decides, and acts. In the open beta it already resolves real incidents end-to-end in under 90 seconds, at under $1 each — and we've demoed it live. Built on the Claude Agent SDK with multi-region failover, available today in KubeBolt Cloud.
When an incident fires, Autopilot launches a long-running investigation session. It correlates events, logs, deploys, and prior incidents — then writes a verdict you can audit.
Proposes a patch, runs it through a deterministic executor with policy guardrails, and rolls back if SLO breaches. You approve once; it remembers.
Generates a draft postmortem from the incident timeline. Action items linked to PRs. Five-whys included. Edit, don't write from scratch.
Schedules power-on / power-off across EKS, AKS, GKE, OpenShift and node pools. Pays for KubeBolt itself in most environments.
You hold the keys
You decide how much autonomy: Suggest only, Approve & execute, or Autonomous (destructive actions always ask for your sign-off). An unattended window with days, hours and timezone, per-org opt-in, and a platform-level autonomy ceiling. All of it, today, in the open beta.
Why Claude exclusively — for Autopilot
For unsupervised action we run Claude exclusively, with multi-region failover across Anthropic API, AWS Bedrock and Google Vertex AI. Autonomous remediation demands the single most reliable model — no router, no model zoo, no ambiguity in a postmortem. Kobi, with you in the loop, is where bring-your-own-model lives.
Module — Cluster Lifecycle
Roadmap · 2026 · Business / EnterpriseMost clusters run 24/7 even though humans don't. KubeBolt's lifecycle module will schedule power-on / power-off across your fleet — preview environments, dev clusters, weekend downtime — and track the savings. Shipping with the KubeBolt Business tier in 2026.
Typical savings
~65%
on non-production cluster spend, when scheduled to nights & weekends with smart warm-up before working hours.
The infrastructure savings often exceed the KubeBolt subscription itself.
Supported platforms
Per-cluster schedules, weekday / weekend rules, manual override, audit log of every power transition. Designed not to interrupt CI runs in flight.
Architecture — Determinism first
All 6 layers ready · Autopilot in open betaA Skill is a declarative, deterministic diagnostic routine — a recipe for a known failure pattern. All six layers are now ready and work together inside Kobi Autopilot, now in open beta: L1 Detectors flags the pattern, L2 Router sets the depth, L3 Investigator finds the root cause, L4 Planner proposes the fix, L5 Executor applies it behind guardrails, and L6 Postmortem writes it up. L1 and L5 also live outside Autopilot today — the 24-rule Insights Engine below and the deterministic write-ops surface above (Set image, Secret reveal with audit, Apply manifest).
L1
READYDetectors
Deterministic. No AI.
L2
READYRouter
Haiku-class. Triage.
L3
READYInvestigator
Sonnet-class. RCA.
L4
READYPlanner
Sonnet / Opus.
L5
READYExecutor
Deterministic. Guardrails.
L6
READYPostmortem
Sonnet / Opus.
Insights Engine
Continuous evaluation against proven heuristics. Actionable recommendations, not raw PromQL. Each rule is a Skill — so the engine works even before any model is configured.
Module — Cluster operations
Available todayEvery kubectl-equivalent verb is a click. Set image, scale, drain a node, edit env, reveal a secret, apply a fresh manifest — all from the dashboard with RBAC enforcement on every endpoint, an audit log entry on every mutation, and a governance switch to scope or turn off write actions. No terminal needed.
Set image, set CPU/memory requests and limits, set env (literal, ConfigMap ref, Secret ref), edit labels and annotations. Strategic-merge patches with init-container support and limit-≥-request guards.
Mandatory operator reason (≥10 chars) before any reveal. Production-namespace gate escalates to Admin. Per-key 60s auto-hide. Audit log records who/what/when/why — never the values.
Cordon, uncordon, and drain with live SSE progress as pods are evicted. Cancel mid-flight. Same toolbar parity from Node detail and the Nodes list.
Paste YAML or pick from 18 starter templates per kind. CodeMirror editor with pre-flight validation: kind / apiVersion / namespace consistency, single-document guard, status / managedFields auto-stripped.
Every rollout's History tab shows a side-by-side YAML diff — what changed, line by line — whether it came from kubectl, GitOps, or KubeBolt. Roll back to either revision in one click.
See every Helm release in the cluster — chart, values, rendered manifest, revision history, and dependencies — decoded straight from Helm's own Secrets. No SDK, no terminal.
Surface — Operator UI
A real-time map of every pod, service, and request path. Mini-dashboards on every resource. Traffic flow you can actually watch. Built for engineers who'd rather see the topology than read another kubectl get.
Pods, Deployments, Services, Ingresses, ConfigMaps, Secrets, Jobs, Nodes — every kind gets a purpose-built view with the right columns, the right actions, the right detail panes.
Every list, graph, and edge is driven by Watch streams over the K8s API. State changes show up in milliseconds — no F5, no stale rows, no surprise.
Type to jump to any resource across any namespace. Fuzzy match across kinds. Recent context pinned. Works the same as your editor — because that's how you already think.
Overview for at-a-glance health, Capacity for right-sizing recommendations and top consumers, Reliability for L7 error rates, hot-spots, and Network Drops via Hubble.
Stack
Source
Kubernetes
Core
KubeBolt Engine
Surfaces
Where you work
Why KubeBolt
Dashboards, metrics and alerts are table stakes — everyone has them. The difference is what happens after the alert fires.
Install
OCI chart on GHCR. Configurable RBAC, Ingress, auth, resources.
helm install kubebolt \
oci://ghcr.io/clm-cloud-solutions/kubebolt/helm/kubebolt
kubectl port-forward svc/kubebolt 3000:80Need the full reference? Read the docs →
Roadmap — By stage
KubeBolt advances in stages: first you see, then you act with approval, finally the platform acts on its own under your policy. Indicative 2026 windows — the stage marks when each capability arrives, not the plan that includes it.
See and resolve
Total visibility
Act with your approval
Autonomy under policy
Plans — One place to choose your path
The open-source agent is free forever under Apache 2.0. KubeBolt Cloud adds hosted Kobi Autopilot, Lifecycle Management and team SSO — with a free hosted tier available now.
Apache 2.0. You operate the infra. KubeBolt operates nothing.
We operate the control plane. You only deploy the agent. Lead-magnet limits.
What's next
KubeBolt Cloud is here, and we're rolling it out in stages — new capabilities, pricing, and more every few weeks. Drop your email and we'll ping you on every release. No noise.
No spam. Unsubscribe with one click. We share product updates only.
The KubeBolt AI SRE · visiting
Kobi can be wrong. Official pricing lives in the plans section.