OpenClaw configuration templates for different LLM providers and
models.
No custom templates. Tenants use the built-in default config.
MCP Tools (via AgentCore Gateway)
All tenant VMs auto-connect to the AgentCore Gateway and gain these
MCP tools. Tool definitions live in
deploy/lambda/agentcore_tools/ +
deploy/stack.py.
no params
AgentCore not enabled, or no tools registered yet. Set
agentcore.enabled: true in config.yml and
redeploy.
Skill Groups
Groups bundle skills together so a tenant can subscribe via
group: "team-sre"
instead of listing every skill. A tenant's effective skill set =
tenant.skills ∪ group.skills. Tenants without scoping
fields get every skill (legacy broadcast).
empty group
No groups defined. Click + New Group to bundle skills for a
team.
Shared Skills
Skills are shared across all tenants. They're plain markdown files in
s3://${ASSETS_BUCKET}/skills/<name>/SKILL.md and
are synced to every host every 5 minutes via cron, then injected into
VMs at launch.
Click a row to view / edit. Use Groups (above) to scope skills to
specific tenants.
💡 To enable: set metrics.enabled: true in
config.yml and redeploy. The stack will provision Amazon
Managed Prometheus + Grafana, and ADOT collectors on each host start
scraping in ~3 minutes after rollout.
Grafana Data Source
In Grafana →
Connections → Data sources → Add data source → Prometheus, fill
in the values below. The Grafana workspace's IAM role already has read
access to AMP, so no static keys are needed.
Prometheus server URL
SigV4 auth
On
Authentication Provider
AWS SDK Default— "Workspace IAM Role" or "Access & secret key" return
403
Default Region
⚠️ Use the server URL above (workspace root) — not the
remote_write URL. Grafana appends
/api/v1/query itself.
ADOT remote_write:
Per-VM Metrics
Each host's host-agent exposes these gauges on
:8899/metrics. ADOT scrapes every 30s and remote-writes
to AMP via SigV4 (no static credentials).
Metric
Type
Labels
Description
openclaw_vm_health
gauge (0/1)
tenant
1 if VM responded to ping, else 0
openclaw_vm_cpu_pct
gauge
tenant
Per-VM CPU usage (percent of allocated vcpus)
openclaw_vm_memory_used_mb
gauge
tenant
Per-VM memory in active use (MB, from VmRSS)
openclaw_vm_memory_balloon_mib
gauge
tenant
Balloon size held by the host (MiB)
openclaw_vm_disk_used_mb
gauge
tenant
Per-VM data disk used (MB)
openclaw_vm_disk_total_mb
gauge
tenant
Per-VM data disk capacity (MB)
openclaw_vm_disk_used_pct
gauge
tenant
Per-VM data disk used (percent)
Sample PromQL
Copy into Grafana → Explore → AMP datasource.
Memory used by all running VMs of a tenant
sum by (tenant) (openclaw_vm_memory_used_mb)
Hosts with at least one unhealthy VM in the last minute
min_over_time(openclaw_vm_health[1m]) == 0
Tenants over 90% disk usage
openclaw_vm_disk_used_pct > 90
💾 Backups
Tenant
Source Status
Backup Time
Size
Actions
↳ previous backup
Backups are retained for 7 days (S3 lifecycle). Orphan backups are from
tenants that have been deleted — restoring creates a new tenant with the
backup's data volume.
🛠 Pool Ops — 集群调度中枢
对整个 openclaw pool 的批量调度行为集中在此,全部走真实控制面端点(无
mock): 批量生命周期(POST /batch/tenants)、滚动重建镜像(POST /hosts/refresh-rootfs)、容量账对账。需 operator/admin 角色。
Must contain at least one top-level # Title line.
Max 256 KiB.
Create Skill Group
You can add/remove skills later via the group's edit
panel.
Source backup
· · orphan
Migrate Tenant
Tenant
· current host
Live migration via Firecracker snapshot/restore. The VM is paused on
the source, snapshotted to S3, restored on the target host, then
resumed. Pause window is on the order of seconds, not
user-perceptible for most agents.
No other active hosts available. Scale ASG out first or wait for
capacity.