io.github.kubeleash/kubeleash

kubeleash

Point it at your over-privileged kubeconfig — it still can't nuke prod.

Guardrails for AI agents on your cluster. kubeleash is a local MCP server for Kubernetes whose differentiator is RBAC-style, context-scoped access control. Point it at a kubeconfig — even a cluster-admin one — and a local policy file constrains what the agent can actually do, per kube context, with destructive actions gated before any call reaches the cluster.

Why

Most Kubernetes MCP servers inherit the kubeconfig's permissions wholesale — whatever the credentials grant, the agent can do. kubeleash adds three things native RBAC can't express for this use case:

• Constrain the agent independently of the credentials. Effective access is always kubeconfig-grants ∩ policy-allows — kubeleash only ever subtracts.
• Context-aware guardrails. Policy varies by context (prod vs staging vs dev); native RBAC is per-cluster.
• Block destructive verbs (delete/exec/…) as a safety net against agent mistakes and prompt injection.

What kubeleash enforces

kubeleash governs only the calls routed through it. It provides a hard guarantee when it is the agent's sole path to the cluster. An agent that also has shell access (raw kubectl, oc, the API) can step around it, so in a shell-enabled assistant kubeleash is advisory — pair it with the using-kubeleash skill, which instructs the agent to stop (not reach for kubectl) when the leash is unavailable.

Policy in 10 seconds

policies:
  - contexts: ".*prod.*"          # regex over the active context name
    allow:
      resources: ["*"]
      verbs: [get, list, watch]   # read-only in prod
    deny:
      verbs: [exec]               # never, regardless of credentials

Deny wins. Default deny. A broken policy refuses to start — it never fails open.

See examples/policy.yaml for a fuller, commented policy (read-only prod, broader staging, namespace-scoped dev).

Quickstart

Install it (Homebrew shown — see Install for go install and the container), then exercise a policy without touching any cluster:

brew install kubeleash/tap/kubeleash

# Grab the commented example policy:
curl -fsSL https://raw.githubusercontent.com/kubeleash/kubeleash/main/examples/policy.yaml -o policy.yaml

# See how kubeleash validates and normalizes a policy (no cluster needed):
kubeleash --policy policy.yaml --print-effective-policy

# Try it without touching any cluster — every decision is logged, nothing runs:
kubeleash --policy policy.yaml --dry-run

Prefer source? git clone https://github.com/kubeleash/kubeleash && cd kubeleash && go build -o kubeleash ./cmd/kubeleash (Go 1.26+).

Then point an MCP client at it (see below). kubeleash speaks MCP over stdio, so it's launched by your client, not run as a daemon.

Install

All channels run kubeleash locally over stdio — your client launches the binary; nothing is hosted. kubeleash is listed in the official MCP Registry as io.github.kubeleash/kubeleash, so registry-aware clients can discover it too.

One-click

The Add to Cursor / VS Code buttons at the top are the fastest path (they need the kubeleash binary on PATH — see Manual). Other clients:

Claude Code — this repo is its own plugin marketplace:

/plugin marketplace add kubeleash/kubeleash
/plugin install kubeleash@kubeleash

No separate install step: on first run the plugin uses a kubeleash already on your PATH, or otherwise downloads the matching release binary (verifying its checksum) and caches it.

Safe by default — no policy authoring required first: if you have no policy at ~/.kubeleash/policy.yaml, the plugin writes a read-only starter there on first run (allow get/list/watch, deny exec/delete). Review and widen it to grant more access. (This convenience is the plugin's only; the raw binary still requires an explicit --policy and refuses to start without one.)

The plugin also ships two skills — using-kubeleash (how an agent should query and act through the gated tools) and authoring-kubeleash-policy (how to write the policy) — so the agent understands the guardrails, not just the tool list.

Claude Desktop — download kubeleash.mcpb from the releases page and double-click it; the bundle ships the binary and prompts you for the policy and kubeconfig.

cursor://anysphere.cursor-deeplink/mcp/install?name=kubeleash&config=eyJjb21tYW5kIjoia3ViZWxlYXNoIiwiYXJncyI6WyItLXBvbGljeSIsIi9hYnNvbHV0ZS9wYXRoL3RvL3BvbGljeS55YW1sIl19

vscode:mcp/install?%7B%22name%22%3A%22kubeleash%22%2C%22command%22%3A%22kubeleash%22%2C%22args%22%3A%5B%22--policy%22%2C%22%24%7Binput%3ApolicyPath%7D%22%5D%2C%22inputs%22%3A%5B%7B%22id%22%3A%22policyPath%22%2C%22type%22%3A%22promptString%22%2C%22description%22%3A%22Path%20to%20your%20kubeleash%20policy.yaml%22%7D%5D%7D

Manual

# Homebrew
brew install kubeleash/tap/kubeleash

# Go
go install github.com/kubeleash/kubeleash/cmd/kubeleash@latest

# Container (great for running the MCP server sandboxed)
docker run --rm -i -v ~/.kube:/root/.kube:ro -v ./policy.yaml:/policy.yaml:ro \
  ghcr.io/kubeleash/kubeleash --policy /policy.yaml

kubeleash runs locally over stdio and talks only to your clusters. There is intentionally no remote/hosted URL connector — it would mean handing your cluster credentials to a third party.

Use it as an MCP server

kubeleash exposes 8 generic, GVK-agnostic tools (k8s_list, k8s_get, k8s_apply, k8s_delete, k8s_logs, k8s_exec, k8s_scale, and k8s_capabilities) that work for any resource, including CRDs. Every call is checked against your policy and recorded to a JSON audit log on stderr (stdout is the MCP transport). It speaks MCP over stdio — point your client at the binary (or container):

// Claude Desktop / Cursor / VS Code MCP config
{
  "mcpServers": {
    "kubeleash": {
      "command": "kubeleash",
      "args": ["--policy", "/absolute/path/to/policy.yaml"],
      "env": { "KUBECONFIG": "/absolute/path/to/kubeconfig" }
    }
  }
}

Privacy

Zero telemetry. No phone-home. Local-only by design. kubeleash talks only to the Kubernetes API servers you point it at — there is intentionally no remote or hosted connector to route your cluster credentials through. That's a feature, not a gap: you run it against real clusters with privileged credentials, so nothing should sit between the agent and your API server but the leash.

Project

• Design · Contributing · Security & threat model · Code of Conduct
• License: Apache-2.0