Kubernetes MCP Server

Community MCP server that wraps kubectl and Helm to let AI agents inspect, deploy, and troubleshoot resources in Kubernetes clusters.

Developer Tools by Flux159 (community) None (uses local kubeconfig) active

GitHub Docs npm: mcp-server-kubernetes

Overview

mcp-server-kubernetes is a community-maintained Model Context Protocol server that connects AI assistants to Kubernetes clusters. It shells out to the local kubectl and helm CLIs using the user's existing kubeconfig, so any context that works in your terminal (minikube, kind, EKS, GKE, AKS, Rancher Desktop, OpenShift, etc.) works through the server.

The server exposes a broad set of kubectl primitives (get, describe, create, apply, delete, logs, scale, patch, rollout, context) alongside higher-level helpers for port-forwarding, pod cleanup (Evicted, CrashLoopBackOff, ImagePullBackOff), node cordon/drain/uncordon, and full Helm 3 chart lifecycle management. A kubectl_generic escape hatch lets the agent run arbitrary kubectl subcommands when no dedicated tool exists.

It is one of the most widely adopted Kubernetes MCP implementations (1.4k+ GitHub stars) and supports a non-destructive mode that blocks deletes and other mutating operations, optional secrets masking, and OpenTelemetry traces. It is distributed as the mcp-server-kubernetes npm package and is typically launched via npx from Claude Desktop, Claude Code, Cursor, or VS Code.

Tools

Tool	Description
`kubectl_get`	Get or list Kubernetes resources of any type.
`kubectl_describe`	Show detailed info about a specific resource.
`kubectl_create`	Create a resource imperatively or from manifest.
`kubectl_apply`	Apply a manifest to the cluster (create or update).
`kubectl_delete`	Delete a resource by type/name or manifest.
`kubectl_logs`	Fetch logs from a pod or container.
`kubectl_scale`	Scale a Deployment, ReplicaSet, or StatefulSet.
`kubectl_patch`	Apply a strategic, merge, or JSON patch to a resource.
`kubectl_rollout`	Manage rollouts (status, restart, undo, pause, resume).
`kubectl_context`	List or switch the active kubeconfig context.
`kubectl_generic`	Run an arbitrary kubectl command for cases not covered by other tools.
`explain_resource`	Run kubectl explain to describe a resource type's schema.
`list_api_resources`	List API resources available on the cluster.
`ping`	Verify connectivity to the cluster's API server.
`port_forward`	Start a kubectl port-forward to a pod or service.
`stop_port_forward`	Stop a running port-forward session.
`install_helm_chart`	Install a Helm chart into the cluster.
`upgrade_helm_chart`	Upgrade an existing Helm release.
`uninstall_helm_chart`	Uninstall a Helm release.
`helm_template_apply`	Render a Helm chart locally and apply the manifests.
`cleanup_pods`	Bulk-clean pods in Evicted, CrashLoopBackOff, ImagePullBackOff, and similar failure states.
`node_management`	Cordon, drain, or uncordon a node.

Setup Guide

Prerequisites

kubectl installed and on PATH
A valid kubeconfig at ~/.kube/config with at least one working context
Access to a Kubernetes cluster (minikube, kind, GKE, EKS, AKS, OpenShift, etc.)
Helm v3 installed (optional, only if you plan to use Helm tools)

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "kubernetes": {
      "command": "npx",
      "args": ["mcp-server-kubernetes"]
    }
  }
}

Claude Code

claude mcp add kubernetes -- npx mcp-server-kubernetes

Cursor / VS Code

Use the same JSON block as Claude Desktop in your MCP-compatible extension's config.

Non-destructive (read-only) mode

Set ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS=true in the server's environment to disable kubectl_delete, uninstall_helm_chart, and other mutating operations while keeping read and create/update access.

Use Cases

Let an AI agent triage cluster incidents by listing pods, pulling logs from CrashLoopBackOff containers, and proposing manifest fixes
Apply or patch Deployment manifests during deploys and watch kubectl_rollout status until ready
Bulk-clean failed pods (Evicted, ImagePullBackOff) across namespaces with a single cleanup_pods call
Install, upgrade, and uninstall Helm releases from chat (e.g. roll out a new ingress-nginx version with custom values)
Cordon and drain nodes safely before maintenance, then uncordon when finished

Example Prompts

"List all pods in the production namespace and show logs for any that aren't Running."
"Scale the checkout Deployment to 5 replicas and watch the rollout until it's healthy."
"Install the bitnami/redis Helm chart as release cache with persistence disabled."
"Cordon node ip-10-0-1-23 and drain it, ignoring DaemonSets."
"Clean up every pod stuck in ImagePullBackOff across all namespaces."

Pros

Broad tool coverage spanning kubectl, Helm, port-forwarding, and node lifecycle in one server
Works with any cluster that already has a working kubeconfig, no extra credentials to wire up
Non-destructive mode and secrets masking make it safer to expose to autonomous agents
Active community project with 1.4k+ stars and frequent releases on npm

Limitations

Community-maintained, not an official Kubernetes project, so support and roadmap are best-effort
Wraps the local kubectl/Helm binaries, so they must be installed and on PATH (no in-cluster Go client)
Powerful by default; without ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, an agent can delete resources or uninstall Helm releases

Alternatives

containers/kubernetes-mcp-server: native Go implementation that talks to the API server directly, with OpenShift, Tekton, Kiali, and KubeVirt tools
alexei-led/k8s-mcp-server: bridges kubectl, Helm, istioctl, and argocd with a focus on safe command execution
rohitg00/kubectl-mcp-server: kubectl-focused MCP server listed in the CNCF Landscape