ironclaw

**Quick judgment** Ironclaw is a useful external safety layer for agent workflows, focused on generic threat classification (prompt injection, secrets, malicious commands, suspicious skill files). It is best treated as an additional filter, not a primary or sole safety mechanism. **Key strengths & benefits** - Custom, criteria-driven classifier: you define what counts as a “threat,” enabling adaptation to your own risk model. - Good coverage of common agent risks (injections, secret leaks, destructive commands, malicious skills) with ready-made criteria templates. - Simple REST API; anonymous access works out of the box for light use, with optional registration for higher limits. - Documentation explains how to write effective criteria, which is crucial for practical performance. **Risks & limitations** - **Classification uncertainty**: The service explicitly notes that it is not 100% accurate; false negatives and false positives are expected. It must not be treated as a strict gatekeeper for safety-critical actions. - **Data exposure**: All checked content is sent to an external service (ironclaw.io). Using it on highly sensitive data or secrets may conflict with your privacy or compliance requirements. - **External dependency**: Relies on a third-party API with rate limits (10/min, 100/day anonymous; 60/min, 10k/month registered; higher via contact). Downtime, quota changes, or network issues will directly affect your safety checks. - **Single-label binary output**: It returns a 0/1 label plus confidence for a single set of criteria; complex multi-class policies require multiple calls or additional logic on your side. **Recommended scenarios** - As a **secondary safety filter** before executing shell commands, installing skills, or forwarding untrusted content to powerful tools. - For **screening untrusted messages and DMs** to reduce prompt-injection attempts reaching your core agent. - For **light- to medium-volume deployments** where sending content to an external classifier is acceptable and rate limits are sufficient. Avoid relying on Ironclaw as the sole protection in high-stakes or highly regulated environments; combine it with local checks, strong internal policies, and human review where appropriate.