OpenRouter's Armor Vets Agent Tool Calls, but Only Flags Them in Claude Code Mode

OpenRouter launched Armor in beta, a safety classifier that checks every agent tool call against plain-English rules and blocks flagged actions. In its Claude Code compatibility mode, however, a flagged tool_use is kept and only marked, leaving the calling application to stop it.

• Armor uses TypeSafe's Jev model and returns a flagged or not_flagged verdict
• Rules come in hard_deny, soft_deny and allow tiers with 0.6 and 0.7 thresholds
• In claude_code_compatibility mode the call is marked in safeguard_results, not blocked
• An audit mode scores calls without blocking anything for rule tuning