Propose an MCP tool call for human approval BEFORE running it. Call this — instead of calling the tool directly — whenever an autonomous or unattended caller (a scheduled standing agent, an unattended agent-runner run, or any MCP client operating without a human watching) is about to perform a write it knows or suspects is risky. The target tool's OWN registered risk hints (readOnlyHint/destructiveHint) decide the tier: GREEN (read-only) tools are never staged — this call is then a no-op passthrough (`result: 'not_required'`) and the caller should just invoke the tool directly. AMBER (reversible write to the caller's own state) and RED (destructive or outward-facing) tools ARE staged: this call does NOT execute anything — it only records the proposal and returns a `staged_action_id`. A human (or any client acting on the human's behalf) later calls `approve_staged_action` or `reject_staged_action` to decide it. Tier: sp500+ (sample rejected — guest has no saved state).
Connector