Getting Started
KubeBolt Cloud
Managed KubeBolt at app.kubebolt.io — nothing to host, with a Free plan and Kobi Autopilot in open beta.
On this page
KubeBolt Cloud is the managed offering of KubeBolt, the Kubernetes operations platform: the same core, run by us at app.kubebolt.io, plus the features that only exist there. You connect clusters with the agent exactly as documented in Connecting Clusters — the agent only ever dials outbound, so no inbound access to your clusters is required.
Sign up at app.kubebolt.io/signup with email or social login (Google, Microsoft, or GitHub); the walkthrough is Getting started on Cloud. Free and Team are available today; Business and Enterprise arrive in stages — see the public roadmap and Pricing.
Open source vs Cloud
| Open source (self-hosted) | Cloud | |
|---|---|---|
| Core platform — dashboards, insights, security findings, actions, multi-cluster | ✅ Apache 2.0, free forever | ✅ |
| Kobi Copilot — you ask, Kobi answers and proposes actions you approve | ✅ With your own LLM key (Anthropic or any OpenAI-compatible API) | ✅ Managed, no key to configure |
| Kobi MCP server (read-only) | ✅ | ✅ |
| Cloud connectors — Kobi reads your cloud account | ✅ AWS and Azure, with your own credentials | ✅ AWS, from the Team plan |
| AI usage | You pay your provider directly | AI credits included per plan |
| Kobi Autopilot — Kobi acts on its own, you approve the destructive part | — | ✅ Open beta, from the Free plan |
| Hosting, upgrades, backups | You | Us |
| Sign-in | Local users and passwords | Email and password, or Google, Microsoft and GitHub |
| Teams & organizations | One organization, a single default team; Viewer / Editor / Admin roles | Organizations and teams with roles |
The open-source core is not a demo — it’s the same engine, including Kobi Copilot, and it stays Apache 2.0. Cloud is for teams that would rather not run it themselves, and for Kobi Autopilot.
Kobi Autopilot (open beta)
Kobi Autopilot is Kobi’s autonomous mode: it investigates an incident and, within the policy your team sets, remediates it. It runs only on Cloud and is in open beta from the Free plan. Its settings live in Administration → AI & Autopilot → Kobi Autopilot. It ships with deliberate guardrails:
- Off until you turn it on — each organization opts in explicitly, and it starts in Suggest only mode: it investigates and recommends, and you apply the change yourself. Approve & execute waits for your approval before every change; Autonomous runs low-risk actions on its own.
- Known incidents only — Autopilot opens an incident only for a closed catalog of incident classes. The deterministic rules surface a problem; they never fix anything on their own.
- A closed set of actions — the agent that executes may only restart,
scale or roll back a Deployment, patch its resources or a probe, or patch an
HPA, plus wait for pods to be Ready or hand the incident to a person. It
never touches namespaces on the blocked list (
kube-system,kube-publicandkube-node-leaseby default, plus any your organization adds). - Verified, or reverted — after a change it waits for the pods to be Ready. If they aren’t, it reverts the change (automatic rollback, on by default) — and only then. When the cause is outside the cluster (a cloud resource, for example), the run ends with a person, never as resolved.
- Unattended time window — you define when it may work on its own, with a real timezone. Outside the window, incidents are parked and visible, not silently worked.
- Destructive actions always require human approval, in every mode.
- Only the plan runs — every change must be one of the run’s planned actions, with the plan’s arguments; anything else is refused.
- What it reads is evidence, never policy — text in annotations, labels, env, logs, cloud tags or External Secrets objects can’t make Autopilot stand down, defer or scale to zero. It records such a claim as unverified and picks the remedy as if it were absent; only your Autopilot settings decide whether to act, which action, the approval and the cooldown.
- When a person must decide, the plan waits — if the investigation says a person must decide, every action that changes something waits for approval, whatever the mode and the risk. The approval (the web card, the queue, Slack, the approval email and webhooks) says when its reason cites the workload’s own text.
- Approving takes a role. Approving or rejecting an action and re-running an incident require Editor; changing Autopilot’s settings requires Admin. Viewers don’t see Approve or Reject. Every decision is recorded with the user who made it, and approvals, re-runs and role rejections are audited.
- One credit pool — Kobi Copilot and Autopilot draw from the same monthly pool of AI credits, and Usage & Credits shows how it splits between them.
Autopilot is a Cloud feature. In the open-source distribution, Kobi Copilot proposes and you approve — see Actions & Governance.
Repeats and credit pacing
Three rules keep Autopilot from spending the month on problems it has already diagnosed. The first and the last are settings in Administration → AI & Autopilot → Kobi Autopilot → Repeats and credit pacing.
- A diagnosis stands. A run that ends in a recommendation (or hands the
incident to a person) leaves a diagnosis that holds for 7 days by default
(Re-investigate an unchanged problem after … days; 0 turns it off). If the
same problem comes back with no material change, Autopilot reopens the
diagnosed incident and counts a recurrence instead of paying for a new run.
A material change is a different rule; for a Deployment’s pods, a different
pod-template-hash(arollout undoback to the same broken spec is not one); otherwise a newgenerationof the owning workload. A Deployment that became available after the diagnosis counts as a change too, so a new cause on the same workload gets a new investigation, but only once the diagnosed incident has closed (its condition stayed gone for the grace period): a readiness probe failing on and off is not a recovery, and while the recommendation or the stop still waits on a person it is the same problem. An incident that waited (for the hourly limit, the schedule window or credit pacing) goes through the same check when the wait clears: if the standing diagnosis covers it, it closes as Covered by another incident instead of running, and does not count against the hourly limit. - One problem, one investigation. A consequence and its cause share one
run whichever fires first: a rollout past its deadline is covered by
image-pull, crash-loop, OOM, missing ConfigMap/Secret and readiness
failures, and a crash-loop by OOM and liveness failures. A StatefulSet’s
replicas (
kafka-0,kafka-1, …) count as one workload. - The month’s credit is spread. With Spread this month’s credits across
the month on (the default), what Autopilot may have spent by a given day is
capped at
pool × (day + days ahead) / days in the month, in UTC. A quiet day leaves its share for later; a busy one can borrow up to 3 days ahead by default. Paced runs go one at a time, and an incident that has to wait is parked asskipped_pacingand picked up again on its own, without a notification. Critical incidents ignore pacing (on by default).
A parked incident (skipped_quota, skipped_schedule, skipped_rate_limit,
skipped_pacing) stays one incident for as long as the problem lasts.
Who hears about a recommendation
A recommendation waits for a person, so Autopilot makes sure one hears about it. In Kobi Autopilot → Who hears about it, Email recommendations and a weekly digest to the organization’s admins is on by default and independent of insight email, so silencing noisy insight mail does not silence Autopilot.
A run that concludes only a person can fix the problem (the cause is outside what Autopilot may change, or the diagnosis is below the confidence floor) ends as a recommendation too, with a warning notice that says Autopilot changed nothing. Failed is kept for runs that tried and did not get there.
- When no notification channel would carry a ready recommendation, it is emailed to the organization’s verified admins who have not opted out: once per incident every 24 hours, at most 20 per organization per day. The email names the organization, the cluster and the workload.
- Once a week, admins get the recommendations nobody has applied yet, grouped by workload across every cluster. With nothing waiting, no email.
- When nothing would reach anyone, Autopilot’s pages say so with a banner (Autopilot can’t notify anyone), and the settings show the delivery status.
Only the organization’s master notifications switch turns both emails off.
Home’s Needs your attention lists what Autopilot waits on a person for: pending approvals, runs that stopped because only a person can go on, and recommendations nobody applied. Mark as handled, on the incident (Editor), closes a recommendation or a stop you took care of or decided not to act on, with an optional note; either also closes itself when its condition clears. See Home and Fleet.
AI credits
Cloud plans include AI credits consumed by Kobi Copilot conversations and Autopilot runs; there is no bring-your-own-key on Cloud. Usage is visible in the unified AI hub, in credits, not dollars: the Copilot and Autopilot usage views, the credit pool, the session and run details and the incident pages all count in the same unit your plan is sold in (Administration → AI & Autopilot → Usage & Credits). The app warns at 80% of the pool, and the organization’s admins get an email as it runs low. When a plan’s monthly credits run out, Kobi pauses until the next month — the non-AI parts of the product keep working.
When Kobi is busy
A Kobi question holds the AI provider for tens of seconds, so each plan sets how many questions can run at once, per organization and per person. A question past that limit waits in line, up to 30 seconds, inside its own stream. The panel shows the wait from the first second: whose limit is full (yours or the organization’s), how many questions are running and how long it has waited. The question starts as soon as a turn frees; past 30 seconds it ends with a message naming the full limit.
The colour says what you have to do:
- Kobi is busy (the wait ran out): blue, the colour of a wait. Ask again in a moment.
- Out of AI credits (the month’s allowance is spent): amber.
- A cloud call held back by the per-minute budget: amber, with a clock, in Kobi and in Autopilot’s run thread. Kobi and Autopilot have separate cloud-call budgets that follow your plan, and a plan change moves them on the next call. See Cloud connectors.
Out-of-date agents by email
Besides the chip on Clusters (Multi-Cluster), the organization’s admins get an email when one of its clusters should upgrade its agent, with the same reasons, command and warnings as the Clusters dialog. Each cluster appears under the name you gave it, with a link that opens its dialog.
- Only update recommended sends mail, once per cluster until a newer release fixes something else: a cluster told about 1.4.9 is not mailed again for 1.4.10.
- The organization’s master notifications switch decides, not the email channel. The mail goes to verified admins who have not opted out, and a cluster counts as told only once someone received it.
Next
Signing up? Go straight to Cloud onboarding. Self-hosting instead? Installation methods covers every way to run KubeBolt yourself.