Kubernetes MCP Server 🔌 MCP server Open source

What it does

Kubernetes MCP Server lets a coding agent inspect and manage Kubernetes and OpenShift clusters. It is written in Go and talks to the Kubernetes API server directly, so it does not wrap kubectl or helm and needs neither installed. It reads your kubeconfig or in-cluster credentials, notices when the config changes, and can work across several clusters at once. For vibe coders, the typical uses are deploying something you just built, reading pod logs and events when a rollout fails, and letting the agent diagnose and fix a broken deployment. It lives in the containers GitHub organisation, the same one that hosts Podman, and is licensed under Apache-2.0.

Tools it gives your agent

• config: list kubeconfig contexts and view the current configuration
• core: CRUD on any resource; pod list, get, delete, logs, top, exec and run; namespaces, events, node logs and stats, and OpenShift projects
• helm (optional): install, list and uninstall releases
• Optional toolsets: Tekton pipelines, KubeVirt, Kiali/Istio, NetObserv and kcp workspaces

Works with

It runs locally over stdio by default. With --port it serves Streamable HTTP, and it supports OAuth/OIDC in that mode. It ships as a native binary, an npm package, a PyPI package and a container image. The docs cover Claude Code, Claude Desktop, VS Code with GitHub Copilot, Cursor and Goose, and any MCP client will work.

How to connect

The project's Claude Code guide uses a dedicated read-only kubeconfig:

claude mcp add-json kubernetes-mcp-server \
  '{"command":"npx","args":["-y","kubernetes-mcp-server@latest","--read-only"],"env":{"KUBECONFIG":"'${HOME}'/.kube/mcp-viewer.kubeconfig"}}' \
  -s user

Cost and limits

The server is free and open source. You need access to a cluster, and any cloud or cluster costs are your own.

Safety notes

The agent acts with whatever rights your kubeconfig grants, including exec into pods and deleting resources. Start with --read-only, or use --disable-destructive to allow creates but block deletes and updates. Point it at a dedicated ServiceAccount, as the project's Kubernetes setup guide recommends, and use denied_resources to hide Secrets. Pick toolsets with --toolsets, and use --disable-multi-cluster to limit it to the current context. The server redacts tokens and keys from the log messages it sends to clients.