Gives Claude three tools to run pre-execution safety checks on autonomous actions through a deterministic certification engine. You call certify_action with scope, reversibility, and sensitivity scores (all 0-1 floats) and get back a decision (Certified, CertifiedWithWarning, EscalateToHuman, or Blocked) plus margin analysis across constraint dimensions. check_budget shows safety budget consumption and rate limits, get_certification_history pulls recent audit logs. Built on QAE's Rust safety kernel with PyO3 bindings. Reach for this when you need reproducible go/no-go decisions for agent actions that affect users or systems, especially if you care about auditing and constraint-based reasoning rather than probabilistic guards.
claude mcp add --transport stdio io.github.tb8412-qae-safety-mcp -- uvx qae-safety