Security auditor & guard-builder for vibecoded apps (web, AI/LLM, mobile, desktop, backend, CI/CD). Claude Code plugin + claude.ai skill. Community project, not an official Anthropic product.
Agent
Use this agent to review the LLM call sites that ClaudeGuardIL's grader enumerated but could not decide by rule — whether an injectable prompt reaches a tool that can actually do damage, whether model output is handled safely, whether the system prompt leaks, and whether the cost path is bounded. Typical triggers…
Agent
Use this agent to adversarially attempt to REFUTE ClaudeGuardIL findings before they reach the report. It has exactly one power — removing a finding by naming the fact that disproves it. It may never raise a finding's confidence, never lower its severity, and never add findings of its own. Typical triggers include the…
Agent
Use this agent to turn a ClaudeGuardIL finding into a concrete, paste-ready fix adapted to the user's real code. Typical triggers include a /cg-harden or /cg-fix request, and any point where a finding needs its guard generated against the project's actual files, framework, and naming. See "When to invoke". Dispatched…
Agent
Use this agent to review the database functions, tables, and infrastructure that ClaudeGuardIL's grader enumerated but could not decide by rule — SECURITY DEFINER bodies that mention auth.uid() without gating on it, RLS policies scoped to the wrong column, and the files in coverage.ungradedSurfaces that no rule walks…
Agent
Use this agent to review the Android and iOS artifacts ClaudeGuardIL's engine enumerated. The grader now decides the definitive manifest facts (debuggable, allowBackup, cleartext, exported components, iOS ATS) and records them in the mobileArtifacts and exportedComponents ledger sets; this auditor takes what those…
Agent
Use this agent to review the web routes, Supabase clients, and Next.js config keys that ClaudeGuardIL's grader enumerated but could not decide by rule. Typical triggers include a /cg-scan run that produced coverage.routes.undeterminable rows, and a request to review whether auth actually enforces, whether ownership is…