KubeGraf

KubeGraf is an AI SRE platform for Kubernetes built by Orkastor, designed to detect, diagnose, and fix cluster incidents autonomously. It combines an in-cluster Go agent with a SaaS control plane and a six-agent AI pipeline to take incidents from detection to remediation in a median of 4 minutes and 21 seconds. The platform is available as both a cloud-hosted SaaS and a self-hosted binary, with an enterprise tier supporting air-gapped deployments.

What It Is

KubeGraf sits above your existing observability stack — Prometheus, Grafana, OpenTelemetry — and acts as an autonomous SRE layer. Rather than surfacing more dashboards and alerts, it reasons over telemetry, proposes evidence-backed remediations, and applies them under a human approval gate. The core product is an in-cluster agent (22 MB, no privileged access required, no inbound ports) that reads cluster state in read-only mode by default, normalizes it into a knowledge graph, and streams allowlisted snapshots to the control plane. Raw secrets and config map values are never transmitted.

The Six-Agent AI Pipeline

The AI engine — branded OrkasAI — runs six specialized agents in sequence when an incident is detected:

• Topology — maps the service graph and blast radius
• RootCause — forms and ranks hypotheses
• LogReasoner — extracts panic lines and error signatures from container logs
• TraceWalker — follows distributed traces across services
• CodeAware — identifies the PR or deploy that introduced the regression
• Remediation — drafts a YAML patch with evidence citations

The pipeline reportedly delivers root cause analysis with 94% confidence and 3–7 evidence citations in a 4–7 second median window.

SafeFix™ and the Approval Model

SafeFix™ is KubeGraf's remediation mechanism. Every patch is HMAC-SHA256 signed end-to-end, dry-run validated against policy and RBAC, and applied to a 10% canary first. If metrics degrade after the canary rollout, the fix auto-reverts within 30 seconds. Approval can be triggered from Slack, the web dashboard, or the CLI — the engineer sees the exact YAML diff and the evidence chain before clicking. The platform positions this as "the right amount of human in the loop" compared to fully autonomous auto-remediation tools.

Deployment Model and Security

KubeGraf ships in two modes. The Cloud product deploys a Helm agent into the cluster that reports to app.kubegraf.io, providing a hosted dashboard, incident management, SLOs, and multi-cluster federation. The Local product is a self-hosted binary where all data stays on the operator's infrastructure. The Enterprise tier ships the full control plane as a Helm chart for on-prem or air-gapped clusters, with tokens and HMAC secrets stored in the customer's KMS. LLM provider is configurable — Anthropic Claude and OpenAI ship by default, with AWS Bedrock and Azure OpenAI available, and enterprise customers can point at a private model endpoint.

Integrations and Platform Support

KubeGraf connects natively to Amazon EKS, Azure AKS, Google GKE, Helm, Prometheus, GitHub, OpenTelemetry, Slack, and ArgoCD. GitOps sync via ArgoCD and Flux is available on higher tiers. PagerDuty and Opsgenie integrations are available for larger teams. The terminal UI and web dashboard provide two interaction surfaces, and the CLI supports one-click approval workflows.

Current Status

KubeGraf is actively developed under Orkastor and is in production availability as of late 2025. The pricing page notes a comparison to Komodor's 2024 removal of their free tier, and the comparison table references publicly available product information dated December 2025. Documentation is versioned under docs-next, signaling active iteration. The platform supports a 14-day free trial on all paid cloud plans with no credit card required.