Security-First Agent Skills
Zero-dependency markdown skills for Claude Code, Cursor, OpenClaw, and any AI agent.
Don't install blind. Guard before you trust.
| Repo | Description |
|---|---|
| Skill Audit | Pre-install security scanner. 7.5% of skills are malicious — audit first |
| Install Guard | 🆕 Pre-install security for npm/pip. Typosquat + supply chain detection |
| Hermes Billing Guard | Detect silent billing routing bugs in Claude Code (HN 984pts) |
| Prompt Guard | Detect & resist prompt injection attacks. 4-layer defense |
| MCP Security Audit | Audit MCP servers before trusting them |
| Model Router | Smart model selection. Route tasks to the right tier, slash costs 70% |
| Token Optimizer | Reduce AI agent token consumption by 98%. Lazy loading + compression |
| Error Doctor | Systematic error recovery. Stop burning tokens on blind retries |
| Session Checkpoint | Save & restore agent context at critical moments |
| Agent Memory | Unified memory layer. Auto-persist, semantic search, cross-session recall |
The numbers:
- 7.5% of 14,706 skills are malicious (RankClaw audit, 2026)
- 59 critical-risk droppers found disguised as legitimate tools (Vett.sh)
- PyTorch Lightning compromised via malicious dependency (Apr 2026)
What malicious installs can do:
- 🔓 Leak your API keys (OpenAI, Anthropic, AWS, Stripe)
- 💀 Execute arbitrary commands on your machine
- 📤 Exfiltrate your source code
- 🔄 Establish persistent backdoors
Package install → Install Guard → (pass) → Install → Prompt Guard (runtime)
↓
(fail) → 🚫 Block
| Layer | Tool | Protects Against |
|---|---|---|
| Package install | skill-install-guard | Typosquat + supply chain |
| Skill install | skill-audit | Malicious skills |
| MCP connect | skill-mcp-security-audit | Malicious MCP servers |
| Runtime | prompt-guard | Prompt injection |
| Billing | skill-hermes-billing-guard | Hidden charges |
GitHub trending + HN proves practical tools win:
mattpocock/skills— 49,600+ stars, +6,187 today. Small composable skills > big frameworksfree-claude-code— 19,200+ stars. Cost control is the #1 pain pointGenericAgent— 8,493 stars. Token efficiency = 6x less consumptionopensre— 4,080 stars. AI agents for incident response- HN 961pts: Claude Code billing bug when commits mention competitors
Our focus: Security foundations + cost control that actually work.
Zero dependencies • Pure markdown • MIT licensed
🔍 Skill Audit • 🛡️ Install Guard • 🚫 Prompt Guard • 💰 Billing Guard