Clusters Overview
An OpenClaw Cluster is a managed VPS running the OpenClaw runtime with multiple AI agents.
What is a cluster?
A cluster is a single Hetzner VPS that runs the OpenClaw gateway with up to 6 agents for $39/mo. Each agent on the cluster gets its own isolated workspace but shares the same server infrastructure.
What’s shared
- Server (CPU, memory, disk)
- LLM model (zai/glm-5)
- Browser, shell, web search capabilities
- Network and bridge sidecar
What’s isolated per agent
- Workspace directory (IDENTITY.md, SOUL.md, USER.md, TOOLS.md)
- Agent directory (auth profiles, sessions)
- Session store (chat history, state)
- Slot ID (unique routing identifier)
Agent-to-Agent Communication
Agents on the same cluster can communicate directly through OpenClaw’s native agentToAgent protocol. This is fast (in-process, no network hop) and enabled by default.
Subagents
Agents can spawn temporary worker subagents to parallelize complex tasks. The cluster config enables this with:
- Max spawn depth: 2 (agent → child → grandchild)
- Max children per agent: 5
- Max concurrent subagents: 8
- Timeout: 15 minutes per subagent
Subagents are ephemeral — they exist only during a single run and disappear after completing their task.
Cluster lifecycle
| Status | Meaning |
|---|---|
| PROVISIONING | Hetzner VPS is being created (3-5 minutes) |
| RUNNING | VPS is up and healthy |
| STOPPED | VPS is powered off |
| ERROR | Something went wrong — check cluster settings |
| DESTROYING | VPS is being deleted |
Health monitoring
Each cluster has a heartbeat chart showing server health over the last 24 hours. The bridge sidecar sends a heartbeat every 60 seconds. Health statuses:
- HEALTHY — heartbeat received, all good
- DEGRADED — heartbeat stale (5+ minutes)
- UNHEALTHY — heartbeat dead (15+ minutes), auto-restart attempted