kubectl-mcp-server is a server based on the Model Context Protocol (MCP), designed to let users manage Kubernetes clusters by conversing with an AI assistant in natural language. It is part of the CNCF Landscape, alongside Terraform.
Purpose:
- Diagnose Pod crashes, optimize costs, deploy applications, audit security, manage Helm Charts, and visualize dashboards through natural language.
- As an AI-driven DevOps assistant, reduce context switching and manage Kubernetes directly within the AI assistant conversation.
Key capabilities:
- 253 powerful MCP tools, covering complete Kubernetes management from Pods to security.
- 8 AI workflow prompts, providing pre-built workflows for common operations.
- 8 MCP resources, browsable Kubernetes data exposure.
- 6 interactive dashboards (HTML UI tools), for visual cluster management.
- 26 browser tools, supporting web automation for cloud providers (EKS, GKE, AKS).
- 107 ecosystem tools, covering GitOps, Cert-Manager, Policy, Backup, KEDA, Cilium, Rollouts, CAPI, KubeVirt, Istio, vCluster, and more.
- Supports multiple transport protocols: stdio, SSE, HTTP, streamable-http.
- Security first: non-destructive mode, secret masking, RBAC validation.
- Advanced diagnostics: AI-driven troubleshooting and cost optimization.
- Multi-cluster support: each tool targets any cluster via the context parameter.
- Complete Helm v3 lifecycle management.
- Powerful CLI, supporting shell-friendly tool discovery and direct invocation.
- Cloud native: can be deployed inside the cluster via kMCP or kagent.
Usage:
- Installation prerequisites: kubectl is installed and configured, with access to a Kubernetes cluster; pip installation requires Python 3.9+, npx installation requires Node.js 14+.
- It is recommended to use npx for zero-install quick start, or use pip installation; Docker is also supported.
- Configure it in the MCP configuration file of an AI assistant (such as Claude Desktop, Cursor, Windsurf, GitHub Copilot, Goose, Gemini CLI, Roo Code, etc.), then restart the assistant.
- Example natural language instructions: "Why is my Pod crashing?", "Deploy a Redis cluster with 3 replicas", "Show which Pods are wasting resources", "Audit the security of all namespaces", "Show the cluster dashboard", etc.
- The built-in CLI can explore and test tools without an AI assistant.
- Supported transport mode options: --transport (stdio, sse, http, streamable-http), --host, --port.
- Security options: --disable-destructive or --non-destructive blocks delete operations, --read-only blocks all write operations.
- Optional features: interactive dashboards (6 UI tools), browser automation (26 tools), kubectl-mcp-app (8 interactive UI dashboards, standalone npm package).
- Enterprise-grade: supports OAuth 2.1 authentication, compatible with OIDC providers such as Okta, Auth0, Keycloak, Microsoft Entra ID, Google OAuth, etc.
- Integration ecosystem: supports Docker MCP Toolkit, agentgateway, kMCP, kagent, etc.
- Agent Skills: provides 25 skill libraries, extending the Kubernetes expertise of AI coding agents, and can use SkillKit to convert formats.
- Multi-cluster: target any cluster via the optional context parameter, without switching contexts.
Applicable scenarios:
- DevOps engineers, SREs, and platform engineers performing daily Kubernetes operations through an AI assistant.
- Troubleshooting: diagnose Pod crashes, network connectivity issues, and resource waste.
- Cost optimization: identify idle resources and provide savings suggestions.
- Security auditing: RBAC permission analysis, secret scanning, Pod security policy checks.
- Application deployment: create Deployments, Services, persistent storage, and health checks.
- Multi-cluster management: cross-cluster operations without switching contexts.
- Enterprise environments: require OAuth 2.1 authentication, audit logs, and multi-tenant support.
Limitations:
- The materials do not provide specific pricing or cost information.
- Some features (such as browser automation and interactive dashboards) are optional and require additional configuration.
- The vCluster (vind) and kind tools run through a local CLI and do not support the context parameter.
- The materials do not mention specific performance metrics or concurrency limits.
