Yuan-lab-LLM/ClawManager
Kubernetes control plane for AI agents that won't let your LLM bill run wild
ClawManager treats AI agent instances like pods you can actually govern—complete with cost controls, skill scanning, and a Redis-backed team bus for multi-agent collaboration.
Velocity · 7d
+21
★ / day
Trend
→steady
star history
What it does ClawManager is a Kubernetes-native platform that spins up, monitors, and governs AI agent runtimes. It wraps three control planes around your cluster: an AI Gateway for audited, cost-tracked model access; an Agent Control Plane for heartbeat-driven runtime orchestration; and a Resource Management layer for reusable skills, channels, and security-scanned bundles. Teams get browser-based workspaces, shared PVC storage, and a Redis Team Bus for multi-agent task dispatch.
The interesting bit
The “Team workspace” MVP is where this gets unusual. Instead of isolated agents, ClawManager provisions member runtimes as pods with injected Redis inboxes, DLQ keys, and shared /team mounts—then keeps DB-backed records so Redis stays a message bus, not your source of truth. It’s basically Slack threads if Slack were a statefulset.
Key highlights
- AI Gateway adds policy-aware routing, audit trails, and cost accounting on top of standard OpenAI-compatible endpoints
- Agent Control Plane handles registration, heartbeat health checks, and desired-state sync for managed runtimes
- Supports OpenClaw (default desktop workspace) and Hermes (Webtop-based
.hermesworkspace) runtimes - Skill Scanner workflows for security review before injection into instances
- One-click Team creation with leader/member orchestration, shared storage, and task dispatch panels
- Standard K8s and K3s deployment manifests included; Go backend, React 19 frontend, MySQL for state
Caveats
- The README is heavy on product tour and light on architecture depth; the developer snapshot trails off mid-sentence (“Deployment a…”)
- Team workspaces are MVP-grade and currently focused on OpenClaw member orchestration; broader runtime support for teams is unclear
- Hermes runtime integration is recent (April 2026); stability at scale is unproven from the sources
Verdict Platform teams running multi-user AI agent infrastructure on Kubernetes should evaluate this. Solo developers or teams without cluster operations in their comfort zone will find the setup cost steep for the payoff.