OpenClaw Kubernetes Operator

As of v0.33.0, the operator provides a hardened environment for OpenClaw fleet management. Key capabilities include:

Security & Hardening

• Default Security Posture: Non-root UID 1000, readOnlyRootFilesystem, dropping ALL capabilities, and seccomp: RuntimeDefault.
• Network Security: Default-deny NetworkPolicy and per-instance RBAC.
• Process Management: Shared PID namespace (enabled by default via spec.shareProcessNamespace: true) for zombie reaping.

Management & Observability

• Lifecycle: Full StatefulSet lifecycle management with auto-update and rollback capabilities.
• Configuration: Support for inline and external config, JSON5, and OpenClawClusterDefaults for singleton cluster-wide settings.
• Observability: OTel Collector sidecar for operator metrics and ServiceMonitor support.
• Storage: PVC management with backup/restore-on-delete and SecretsReady conditions.

Agent Capabilities

• Skills & Plugins: Declarative skill installation and declarative plugin installation via spec.plugins.
• Sidecars: Support for Chromium (with anti-bot CDP proxy), Tailscale (with persistent state), ttyd web terminal, and Nginx gateway proxy.
• Self-Configuration: OpenClawSelfConfig allows agents to perform self-modifications based on an allowlist policy.

OpenClaw agents can autonomously modify their own environment (installing skills, patching config, adding env vars, or seeding workspace files) by interacting with the Kubernetes API.

To enable this capability, you must explicitly allow it in the OpenClawInstance specification using the selfConfigure field. Once enabled, the agent can create OpenClawSelfConfig resources to request changes. The operator validates every request against an allowlist policy to ensure protected configuration keys are not overwritten and unauthorized actions are denied.

# 1. Enable self-configure on the instance
spec:
  selfConfigure:
    enabled: true
    allowedActions: [skills, config, envVars, workspaceFiles]

# 2. The agent creates this to install a skill at runtime
apiVersion: openclaw.rocks/v1alpha1
kind: OpenClawSelfConfig
metadata:
  name: add-fetch-skill
spec:
  instanceRef: my-agent
  addSkills:
    - "@anthropic/mcp-server-fetch"

The OpenClaw operator automatically manages several configurations and settings without requiring explicit user setup:

The SelfConfig controller utilizes Kubernetes Server-Side Apply (SSA) with the field manager name openclaw-selfconfig. This allows for granular ownership of specific fields, preventing accidental overwrites of configuration managed by other entities.

Ownership Granularity

• Skills (+listType=set): Each skill name is individually owned. Multiple field managers can own different skills on the same instance.
• Env vars (+listType=map, key: name): Each environment variable is individually owned by the field manager that last set it.
• Workspace files (map fields): Each file entry under initialFiles is individually owned.
• Config raw: This is owned atomically as a single field.

Handling Conflicts

If a SelfConfig request attempts to remove an item owned by a different field manager, the removal is skipped. In this case, the operator emits a Warning event with the code SelfConfigSkippedRemoval, and the resource status message will include details about the skipped removal.

The OpenClawSelfConfig Custom Resource Definition (CRD) allows a running agent to request modifications to its own OpenClawInstance specification. This process is governed by an allowlist policy to ensure security and stability.

Lifecycle

1. Creation: An agent creates an OpenClawSelfConfig resource. The status is initially Pending.
2. Validation: The operator checks the parent OpenClawInstance for:
   • selfConfigure.enabled must be true (if false, status becomes Denied).
   • Requested action categories must be present in allowedActions (if not, status becomes Denied).
   • Protected config keys (e.g., gateway.*) or restricted environment variables are requested (if so, status becomes Failed).
3. Application: The operator applies the requested changes to the parent instance spec.
4. Completion: Status transitions to Applied (success) or Failed (error).
5. Cleanup: An owner reference is set to the parent instance for garbage collection, and terminal requests are automatically deleted after 1 hour.

apiVersion: openclaw.rocks/v1alpha1
kind: OpenClawSelfConfig
metadata:
  name: add-fetch-skill
spec:
  instanceRef: my-agent
  addSkills:
    - "mcp-server-fetch"
  addEnvVars:
    - name: MY_CUSTOM_VAR
      value: "hello"