Kubernetes MCP Server

Kubernetes MCP Server is for teams that want an AI agent to observe or operate Kubernetes and OpenShift. It is a native Go implementation that talks directly to the Kubernetes API without requiring kubectl or the Helm CLI at runtime, and it ships through local stdio, containerized Streamable HTTP, npm, PyPI, and native binaries. The project organizes generic resources, pods, events, Helm, and optional ecosystem integrations as MCP tools. That usefulness comes with a high-trust boundary: the default configuration is not read-only. A safer starting point is a dedicated ServiceAccount, least-privilege RBAC, a separate kubeconfig, and the --read-only flag, followed by narrowly reviewed write access only when a concrete workflow requires it.

Create a dedicated read-only identity and kubeconfig, then register the npx package with Claude Code using --read-only. Verify the connection with a side-effect-free namespace query.

Install Kubernetes MCP Server from https://github.com/containers/kubernetes-mcp-server. First read the README, docs/getting-started-claude-code.md, docs/getting-started-kubernetes.md, and docs/configuration.md. Do not use my everyday administrator kubeconfig. Confirm the target cluster and namespace, then create a dedicated ServiceAccount, least-privilege read-only RBAC, and a separate kubeconfig. Connect it to Claude Code through npx with --read-only. Ask before requesting credentials, adding permissions, overwriting files, exposing a network endpoint, or running any write operation. Finally run claude mcp list and one side-effect-free namespace listing. Report the config location, identity, permission scope, and result without printing any token.

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

1. 1
   Create a least-privilege identity

   Follow the Kubernetes guide to create a dedicated namespace and ServiceAccount, for example kubectl create namespace mcp and kubectl create serviceaccount mcp-viewer -n mcp. Prefer a namespace-scoped view RoleBinding, using a ClusterRoleBinding only for an explicit cross-namespace read requirement.

2. 2
   Generate a dedicated kubeconfig

   Use docs/getting-started-kubernetes.md to obtain a short-lived ServiceAccount token and write the API server, CA, token, and dedicated context to a separate kubeconfig. Do not reuse an everyday administrator configuration or print the token in logs.

3. 3
   Register the server in read-only mode

   Run claude mcp add-json kubernetes-mcp-server '{"command":"npx","args":["-y","kubernetes-mcp-server@latest","--read-only"],"env":{"KUBECONFIG":"/absolute/path/to/mcp-viewer.kubeconfig"}}' -s user after replacing the kubeconfig path.

4. 4
   Run a side-effect-free verification

   Run claude mcp list and check for Connected, then ask the client to list namespaces. Confirm the target cluster, identity, and returned scope, with no create, update, delete, or exec action.

5. 5
   Narrow any later write access

   If a later workflow genuinely needs mutations, define toolsets, enabled_tools, disabled_tools, denied_resources, and confirmation_fallback=deny before adding the minimum RBAC for one operation.

Based on the public implementation and documentation, the project combines broad Kubernetes coverage and flexible distribution with a substantial set of controls for read-only access, resource filtering, OAuth/OIDC, and observability. The critical caveat is that the default mode is write-capable and confirmation fallback permits actions when a client cannot elicit confirmation. Treat it as an operations interface that requires deliberate RBAC design, never as something to attach directly to production with an administrator kubeconfig.