K8sGPT exists to shorten the distance between a broken cluster and a fix. Instead of grepping through events and hoping for a pattern, you point the CLI at a kubeconfig and run an analyze command. Roughly two dozen built-in analyzers encode common SRE knowledge — problems with Pods, Services, PVCs, Ingress, HPAs, webhooks, network policies and more — and flag what looks wrong.
Because analyzers are rule-based, a scan is useful even without any AI. The optional step is explanation: add a backend such as OpenAI, Azure OpenAI, Amazon Bedrock, Vertex AI, Gemini, Groq, Cohere, watsonx, Hugging Face, Ollama, LocalAI or a custom REST endpoint, and K8sGPT translates the raw output into a diagnosis with recommended actions. Sensitive data can be anonymized before it leaves the cluster, and analysis can be run fully locally with Ollama or LocalAI.
The operator is where the project gets interesting for platform teams. A single custom resource turns K8sGPT into a control loop that continuously watches workloads, records findings, sends notifications to tools like Slack, and — when you explicitly enable it and set a risk threshold — applies fixes automatically. Teams can extend coverage with custom analyzers written in any language over gRPC, or pull in findings from Trivy, Prometheus, Kyverno, KEDA and AWS.
