ClawManager is a Kubernetes-native AI agent instance management control plane that provides governed AI access, runtime orchestration, and reusable resources across multiple agent runtimes.
Purpose ClawManager brings AI agent operations into Kubernetes, integrating managed runtimes, collaborative teams, governed model access, reusable resources, a skills center, and platform security. Users work through a browser-based desktop, while administrators retain visibility and policy control without needing to expose Kubernetes details. It is aimed at the following people: platform teams that run AI agent instances for multiple users; operations personnel who need runtime visibility, command distribution, and desired-state control; and builders who want governed AI access and reusable resource injection rather than manual configuration instance by instance.
Main capabilities Runtime and instance management: Create OpenClaw, Hermes, OpenCode, or DeepSeek Harness workspaces, choose Lite or Pro mode, enabled system images, resource presets or custom CPU/memory/storage, and manage lifecycle, desktop access, files, Shell access, environment variables, archiving, share links, and Lite batch operations in one place.
Northbound API and enterprise access: Integrate with external systems through a separately deployed northbound gateway, supporting one-time JWE authentication, scoped tokens, owner isolation, idempotent asynchronous operations, rate limiting, audit records, and gateway-to-core mTLS. Administrators can also configure LDAP/LDAPS or StartTLS authentication, preview and import directory users, and map groups to platform roles.
AI Gateway: A governed model entry point for managed runtimes, covering five user-facing areas: models, AI audit, cost, session usage, and risk rules. Supports OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages compatibility; persistent model configuration, provider health, pricing, and managed Thinking where supported; secure model routing for sensitive policies; end-to-end audit and trace records; built-in cost accounting and token usage by instance and by session; risk control rules that can intercept or reroute requests.
Agent Control Plane: The runtime orchestration layer for managed AI agent instances, enabling each instance to register, report status, and receive commandsA managed runtime that keeps instances aligned with platform-side desired state. It includes Agent registration with secure bootstrapping and session lifecycle, heartbeat-driven runtime status and health reporting, desired-state synchronization between the control plane and instances, runtime command dispatch for start/stop/config apply/health check/skill operations, and instance-level visibility into Agent status, channels, skills, and command history.
Resource management: A user-side OpenClaw configuration center for reusable startup content, divided into resources, resource packages, and injection records, independent from the administrator security protection console. It includes channel configuration with templates, form/JSON editing, cloning, and lifecycle control; skill ZIP import, conflict handling, download, and deletion, with the skill center additionally providing catalog, ownership, versions, publishing, and installation; simple and advanced editing modes for scheduled tasks; creation, editing, cloning, and reuse of resource packages during instance creation; and a read-only injection snapshot display showing mode, resources, environment variables, status, and creation time.
Team collaboration: Adopts a Leader-mediated workflow. Teams can be created from eight immutable built-in templates or user-owned custom templates, and then the goal is described in team chat. The OpenClaw Lite Leader is responsible for planning and decomposing requests, distributing work, verifying member deliverables, and publishing unified results. Each team has one OpenClaw Lite Leader; each Worker can use OpenClaw Lite or Hermes Lite when available. Custom teams can be generated from natural language intent, refined by role, regenerated, and reused. Team chat records plans, assignments, progress, reviews, deliverables, and final summaries; the execution board tracks the current query, task decomposition, and member delivery status; shared files and artifacts retain collaborative output.
Security protection platform: An independent administrator workspace with a real-time overview integrating four alert metrics, recent events across products, Pod Live Aegis configuration, report export, and emergency circuit breaker. The overview presents the KSecure model as seven risk surfaces, fifteen defense scenarios, and four defense layers. The drill-down areas cover runtime defense, host/container isolation, component trust, egress and identity governance, policy, collaboration, quotas, approvals, skill scanner, and full-chain audit.
Usage During deployment, distinguish between Kubernetes distributions and storage configurations: first choose k3s or k8s, then choose the storage configuration that matches the cluster form. The provided deployment files include deployments/k3s/single-node/clawmanager.yaml, deployments/k3s/cluster/clawmanager.yaml, deployments/k8s/single-node/clawmanager.yaml, and deployments/k8s/cluster/clawmanager.yaml. The cluster configuration is validated using Longhorn (longhorn for RWO data, longhorn-rwx for RWX workspaces), but these StorageClass names are examples only and can be replaced with any CSI class that provides the same access modes. An operations-oriented quick start and first login flow can be found in the user guide, while deployment instructions and architecture-level background can be found in the deployment guide.
Lite mode provisions OpenClaw, Hermes, OpenCode, and DeepSeek Harness instances through a shared gateway runtime pool. Each workspace runs as an isolated gateway process within a managed runtime Pod, starts quickly, and reduces dedicated CPU, memory, storage, and GPU allocation overhead while retaining workspace access, share link/password access, supported channel and skill injection, and administrator visibility. Pro mode provisions a dedicated desktop runtime for each instance, backed by an independent Kubernetes Deployment, Service, and PVC, making it suitable for scenarios that require stronger isolation, full desktop resources, runtime events, instance skill management, and a complete desktop management experience.
Use cases Suitable for platform teams running AI agent instances for multiple users, operations personnel needing runtime visibility and command distribution, builders wanting governed AI access and reusable resource injection, and enterprise environments requiring team collaboration, model governance, cost auditing, and security protection.
Limitations The material does not provide specific information such as pricing, charging methods, performance data, or an explicit platform compatibility list.
