MCP Server Kubernetes
Not yet source-checkedMCP server that lets Claude, Cursor, VS Code and other agents drive kubectl and Helm against real clusters
Machine-drafted from the vendor's site · not yet source-checkedProfile updated Link checked How we rate
In short
This MCP server allows AI clients like Claude and Cursor to execute kubectl and Helm commands against live Kubernetes clusters. It is best for platform engineers who want natural-language control over cluster operations and troubleshooting.
Platform, SRE and DevOps engineers who already live in kubectl and want an MCP-driven Claude, Cursor or VS Code copilot to triage pods, drive rollouts and run Helm operations against clusters they already have kubeconfig access to.
Teams that need a hardened, multi-tenant control plane with approvals, audit and RBAC scoping enforced by the tool itself, or non-Kubernetes shops — this is a thin, trust-your-kubeconfig bridge, not a managed platform.
MCP Server Kubernetes (mcp-server-kubernetes) by Flux159 is a Model Context Protocol server that exposes a kubectl- and Helm-shaped toolset to any MCP-compatible client, so an AI assistant like Claude Desktop, Claude Code, Cursor, VS Code Copilot or Codex CLI can inspect and manage a real Kubernetes cluster through natural-language prompts. It reads the local kubeconfig, so whatever context and RBAC the operator already has, the model gets. Once wired up, the assistant can get / list / describe / create / apply / delete / patch arbitrary resources, tail and grep pod logs, exec commands inside containers, drive rollouts and scale deployments, cordon and drain nodes, port-forward services, and clean up pods stuck in bad phases. Helm v3 is a first-class citizen: install, upgrade, uninstall and template charts with custom values files. A built-in /k8s-diagnose prompt walks the model through a systematic pod-troubleshooting flow (events, describe, logs, previous logs, related resources), which is where the project really earns its keep for on-call and platform engineers. For safety the server ships a non-destructive mode that hides mutating tools, plus secrets masking so credential values do not leak back into the chat transcript, and optional OpenTelemetry tracing so every tool call the model makes is auditable. Installation is a single npx invocation, distributed as a TypeScript/Node package on npm, and the codebase is MIT-licensed on GitHub with an active release cadence and a large fork community. It suits SREs, platform teams and Kubernetes-heavy developers who want an LLM copilot for day-2 operations without building their own tooling layer, and who are comfortable trusting an agent with cluster credentials at the permission level the kubeconfig grants.
This is the reference implementation most people think of when they say 'give Claude kubectl'. The tool coverage is genuinely broad, the diagnose prompt is a nice touch, and non-destructive mode plus secrets masking show the author thought about the obvious footguns. Just remember it inherits your kubeconfig's power — point it at a cluster where an over-confident agent can safely be wrong.
— The AI Tool Bible editorial team
Pros
- ✅ Broad kubectl surface — CRUD on any resource, logs, exec, port-forward, rollouts, scale, cordon/drain and stuck-pod cleanup out of the box
- ✅ First-class Helm v3 support including install, upgrade, uninstall and template with custom values
- ✅ Built-in /k8s-diagnose prompt gives the model a structured troubleshooting recipe rather than free-styling kubectl
- ✅ Non-destructive read-only mode and automatic secrets masking reduce the blast radius of an agent going off-script
- ✅ Works with the mainstream MCP clients (Claude Desktop, Claude Code, Cursor, VS Code, Codex CLI) via a single npx command
- ✅ Optional OpenTelemetry tracing on every tool call, so agent actions are auditable in your existing observability stack
- ✅ MIT-licensed, TypeScript, actively released (v4.x line in 2026) with 1.5k+ stars and 270+ forks
Cons
- ⚠️ Runs with whatever RBAC your kubeconfig has — a compromised or over-eager agent can delete production resources unless you enable non-destructive mode
- ⚠️ No native multi-cluster policy layer or approval workflow; safety relies on kubeconfig scoping and the client's tool-approval UI
- ⚠️ Requires kubectl (and Helm for chart operations) preinstalled and on PATH — not a zero-dependency install
- ⚠️ Node.js / npx runtime is another moving part to manage on operator workstations or bastion hosts
- ⚠️ Advanced cluster features (custom operators, service meshes, cloud-provider APIs) are only reachable through generic apply/patch, not first-class tools
Facts
- Open source
- Yes (license: MIT)
- GitHub stars
- 1,610 (checked Oct 29, 2026)
- Last GitHub push
- Oct 26, 2026
- First commit
- Dec 2024
Where this tool sits
Use cases
Frequently asked
- Which AI clients are compatible with MCP Server Kubernetes?
- It works with MCP-compatible clients including Claude Desktop, Claude Code, Cursor, VS Code Copilot, and Codex CLI.
- Does the server support Helm operations?
- Yes, Helm v3 is a first-class citizen, supporting install, upgrade, uninstall, and template charts with custom values files.
- How does the tool handle safety and permissions?
- It reads the local kubeconfig to inherit existing RBAC permissions. It includes a non-destructive mode to hide mutating tools and automatic secrets masking to prevent credential leaks.
- What is the licensing and cost of the tool?
- The tool is free and open source under the MIT license. Users are only responsible for separate billing of their Kubernetes cluster and AI client subscriptions.
- What are the system requirements for installation?
- Installation requires a single npx invocation. It requires kubectl and Helm to be preinstalled and on the PATH, along with a Node.js runtime.