Your AI assistant finally learns kubectl

What it does

This MCP server is essentially a disciplined wrapper around your local kubectl and helm binaries. It exposes cluster operations—inspecting resources, applying manifests, streaming logs, scaling deployments, and running Helm charts—as structured tools that any MCP-compatible AI assistant can invoke. Instead of replacing the Kubernetes CLI, it turns it into an API that an LLM can call conversationally.

The interesting bit

The project leans into the fact that kubectl is already the de facto control plane. It adds a k8s-diagnose prompt that walks the model through systematic pod troubleshooting, and a non-destructive mode that disables delete operations so a chatbot can’t accidentally evict your production namespace. There is also opt-in OpenTelemetry tracing, which lets you audit exactly which commands the AI executed and how long they took—useful when you are letting an LLM touch real infrastructure.

Key highlights

• Wraps core kubectl workflows (kubectl_get, apply, logs, scale, patch, rollout) and Helm lifecycle operations (install, upgrade, template-based uninstall).
• Includes cluster maintenance utilities: port forwarding, node cordoning/draining, and automated cleanup of stuck pods in states like Evicted or CrashLoopBackOff.
• Masks sensitive data in kubectl get secrets output by default.
• Offers a read-only / non-destructive mode via environment variable for safer AI experimentation.
• Optional distributed tracing exports to any OTLP backend (Jaeger, Tempo, Grafana, etc.).

Caveats

• Secrets masking applies only to kubectl get secrets and explicitly does not affect logs, so sensitive data may still appear in log output.
• The project is a thin wrapper around local binaries: you must already have kubectl and Helm v3 installed and configured on the host.
• Badge links in the README still point to yourusername placeholders, suggesting some repo metadata remains unfinished.

Verdict

Worth a look if you want an AI assistant to handle routine Kubernetes toil—checking pod status, tailing logs, or debugging crashes—without giving it unrestricted shell access. Skip it if you need a deeply embedded Go client, complex multi-cluster federation logic, or an AI that can reason about Kubernetes without relying on external CLI binaries.