k8s-readonly-mcp

A read-only Model Context Protocol (MCP) server that lets an LLM inspect a Kubernetes cluster — list pods, read logs, describe resources — but never mutate it.

Why I built it

I spend a lot of time embedded in customer Kubernetes environments. The most common thing I want from an LLM is "help me understand what's happening in this cluster" — without ever giving it the ability to change anything. So I built the safety in structurally rather than trusting the model to behave.

Every tool call routes through a single function that checks the kubectl verb against an allow-list of read-only verbs (get, describe, logs, top, …). There is no code path that can apply, delete, scale, or exec. If the model asks for a mutation, the server refuses. This is the same least-privilege, human-in-the-loop instinct that any production agent needs.

Related MCP server: kubernetes-mcp

What it does

Quick start

# 1. Install (using uv — https://docs.astral.sh/uv/)
uv sync

# 2. Make sure kubectl points at a cluster.
#    A local cluster is perfect for trying this safely:
#    kind create cluster   (or: minikube start)
kubectl get nodes

# 3. Run the server
uv run k8s-readonly-mcp

Connect it to Claude Desktop

Add this to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "k8s-readonly": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/k8s-readonly-mcp", "run", "k8s-readonly-mcp"]
    }
  }
}

Then ask Claude things like "What pods are failing in the default namespace, and why?" — it will call list_pods and describe_pod, reason over the output, and explain — but it physically cannot change your cluster.

Design decisions

• Allow-list, not block-list. I enumerate what's permitted rather than trying to block bad verbs. A block-list is one forgotten verb away from a mutation; an allow-list fails closed.

• One chokepoint. All kubectl execution goes through _run_kubectl. The security boundary is one function you can audit in 30 seconds.

• Timeouts and clear errors. Calls time out and surface kubectl's stderr instead of hanging or failing silently — the model gets actionable feedback.

What I'd do next

• Add resource-level scoping (restrict to specific namespaces per connection).

• Stream large log outputs instead of buffering.

• Add a small eval that checks the server refuses every mutating verb.

License

MIT