K8sGPT scans a Kubernetes cluster for broken resources and explains what is wrong in plain English. Under the hood it is a Go CLI (and a companion operator) that runs a set of hard-coded “analyzers” — one per resource type — which look for well-known failure modes: pods in ImagePullBackOff, services with no matching endpoints, ingresses pointing at nothing, RBAC roles with dangling subjects, and so on. The analyzer output is already useful on its own (k8sgpt analyze prints a structured list), and then optionally an LLM is called to translate each finding into a remediation suggestion.
The LLM layer is pluggable: OpenAI, Azure OpenAI, AWS Bedrock, Cohere, Google Vertex AI, Ollama for local models, and several others. Sensitive fields can be anonymized before they leave the cluster. The operator variant (k8sgpt-operator) runs the same analyzers continuously and emits Result CRs that can be surfaced in Grafana or Backstage, so the tool becomes an always-on triage layer instead of an interactive CLI.
It was started by Alex Jones and donated to the CNCF sandbox in late 2023. It is one of the cleaner examples of LLM-assisted ops because most of the value comes from the deterministic analyzers; the model is just doing natural-language rewriting on top.