Safety, Costs, Backup, and Recovery

Safety is layered

RapidClaw does not rely on one prompt or one approval checkbox. Identity, connector scope, discovered-tool review, Permission Rules, agent instructions, action policies, approvals, audit, and Emergency Stop work together around the OpenClaw runtime.

• Connector and tool access is deny-by-default and must be explicitly reviewed.
• Writes, external communications, enablement, destructive actions, and other high-impact operations require the applicable explicit approval.
• Schedules begin disabled unless activation is intentionally approved.
• OAuth, entitlement, connector, and tenant checks fail closed when the required state cannot be verified.
• Activity Log, Agent Journal, and approval history retain operational evidence for review.

Emergency Stop

Emergency Stop freezes governed agent execution, scheduled tasks, and Microsoft Teams activity. Arcy remains available so an administrator can diagnose the problem and work through recovery without restoring live behavior first. On mobile, Emergency Stop and Arcy are the intentionally available Command Center controls.

Models, usage, and costs

• Setup deploys a verified direct-model plan from the compatible Microsoft Foundry / Azure AI models available in the selected region and subscription.
• Foundry Model Router availability may be detected, but it is not the default deployed runtime path.
• Models shows the live deployed inventory and supports per-agent assignment.
• Usage & Costs reports every agent, including Arcy, and supports projected-spending review even when an agent currently has zero usage.
• The wizard configures an Azure cost alert using the budget and email supplied by the administrator.
• Azure infrastructure and model-token charges are separate from the RapidClaw product entitlement and are billed through Azure or the configured model provider.

Backup and migration

Use Backup & Migration before major configuration changes, recovery work, or a planned move. It is the administrator entry point for preserving and restoring the OpenClaw workspace. Backups are encrypted and remain part of the governed customer deployment lifecycle.

1. Review current runtime health and resolve active failures before taking a planned backup.
2. Record the deployed product, region, model inventory, active connectors, and important schedules.
3. Create the backup from Command Center and wait for its completion status.
4. Keep the recovery material within the approved customer security boundary.
5. After a restore or migration, verify runtime health, connectors, models, Teams, email, schedules, approvals, and a controlled agent test before broader use.

Updates and recovery

RapidClaw installs the latest release from the verified OpenClaw extended-stable channel at deployment time. Updates use pinned, checksum-verified artifacts and retriable server-side steps. This is a managed upgrade path, not an unattended promise that every upstream release is applied automatically.

• Use Setup to reconnect to an existing authorized deployment rather than creating a second runtime.
• Use Arcy, Dashboard, and the relevant readiness pages to identify whether a failure is in runtime health, models, connectors, Teams, email, entitlement, or configuration.
• Back up before major updates or recovery operations when the current workspace is still accessible.
• Keep schedules and sensitive tools disabled until post-update verification is complete.

Reset and teardown

Reset is manifest-driven and checks every resource category in dependency order. It removes product-created email resources, Teams artifacts, Azure runtime resources, deployment state, connector authorization owned by the deployment, and the customer application registration where applicable. Missing optional resources are skipped rather than treated as failures.