cashclaw-guard

What problem does it solve?

CashClaw Guard stops unattended AI agents from blowing through budgets via runaway loops and getting stuck in infinite recursion that keeps calling the LLM or tools.

Core Features & Use Cases

• Hard cost caps with real-time enforcement: Enforces per-call and per-day USD limits (fails fast instead of waiting for soft limits).
• Recursive call detection with kill switch: Detects repeating prompt/tool fingerprints within a time window and terminates when thresholds are hit.
• Tool firewall and rate limiting: Denies unauthorized tools (e.g., shell/exec/eval) and applies per-tool rate caps to reduce blast radius.
• Operational alerts: Dispatches notifications (Telegram/Slack/Discord per configuration) when budget, recursion, or tool-denial events occur.
• Use case: Run a production cron agent that periodically checks work and must never exceed a strict spend cap; if recursion starts, Guard kills the loop and alerts you before large charges accumulate.

Quick Start

Ask your agent to run CashClaw Guard initialization and test it once so the policy is written, then verify the guard behavior with a dry-run scenario set.