An SRE in your shell, minus the on-call pager

What it does

k8sgpt scans Kubernetes clusters and translates opaque failures into plain English. It runs built-in analyzers across resources like pods, nodes, ingresses, and deployments, then pipes the relevant context to an LLM of your choice. You can run it as a one-off CLI diagnostic or deploy it continuously inside the cluster via its operator, where it integrates with Prometheus and Alertmanager.

The interesting bit

The project treats SRE expertise as data rather than folklore: the analyzers encode operational heuristics for what actually matters in a failure, stripping noise before the AI ever sees it. It also speaks MCP (Model Context Protocol), so Claude Desktop can query cluster health as a native tool rather than via copy-pasted terminal output.

Key highlights

• Built-in analyzers cover core resources like Pods, PVCs, Services, Deployments, Nodes, and Ingresses, with optional filters for Gateway classes, HTTP routes, network policies, logs, storage, a security analyzer, and operator resources such as CatalogSources and ClusterServiceVersions.
• Supports OpenAI, Azure, Cohere, Amazon Bedrock, Google Gemini, and local models out of the box.
• Includes an anonymization flag to scrub sensitive names before sending them to external LLMs.
• Can surface official Kubernetes documentation alongside AI explanations via the --with-doc flag.
• Offers a gRPC serve mode and an MCP server for integration with external clients like Claude Desktop.

Caveats

• The default AI provider is OpenAI, so avoiding external API keys requires explicitly switching to another backend or a local model.
• Homebrew installation on WSL or Linux currently fails with missing build tools and falls back to compiling from source instead of using prebuilt bottles.

Verdict

Platform engineers and SREs who spend too much time decoding raw cluster events will find this a useful filter. If your clusters are already perfectly observable or you refuse to let LLMs near your resource names, look elsewhere.