What
Adding Aufgaard to the Development & Code Tools section. Aufgaard is a Claude Code (and Cursor) plugin that bundles 7 production-AI MCPs + 13 skills + 6 hooks + 5 monitors into one installable — an operator cockpit for production AI.
Repo + license
Real use case
Aufgaard catches the failure-pattern classes that are well-documented in the production-AI operator community:
- Destructive bash chains (the chiefofautism / xornullvoid
rm -rf and glob-wipeout cases) — bash-vet's 30-rule scanner with chain-mode severity escalation
- Cost spikes / 429-imminent (HERMES.md billing-routing surprise) — cost-tracker's
predict_429_in_window
- Agent-config trust-boundary attacks (Cursor CVE-2026-26268, Gemini CLI CVSS 10.0) — skill-vetter v1.1's
vet_agent_config
- Output-action divergence (chiefofautism viral case + Codex sandbox-escalation) — output-vetter's
verify_action_outcome
- Silent cron failures (exit-0 with empty output) — silentwatch
- Pre-upgrade regression risk — upgrade-orchestrator's regression catalogue + provider-side fingerprint
- Deployment health — health-mcp's vital-signs panel
What's in the plugin
- 7 MCP servers bundled via
.mcp.json
- 15 skills (10 invocable + 5 reference):
production-audit (marquee, calls all 7 MCPs in parallel), vet-bash, vet-skill, vet-config, cost-overview, health-check, should-i-upgrade, silent-failures, verify-claim, find-swallowed-exceptions + 5 reference skills (production-patterns, mcp-tool-quick-reference, failure-mode-glossary, operator-cheatsheet, audit-rubric)
- 6 hook bindings (PreToolUse(Bash) → bash-vet auto-block, PreToolUse(agent-config writes) → vet-config auto-block, PostToolUse(Task) → verify-claim, SessionStart → daily overview, etc.)
- 5 specialized subagents (
production-auditor, bash-safety-officer, skill-vetter-officer, cost-optimizer, upgrade-advisor)
- 5 background monitors for proactive notifications
- Cross-platform manifest (Claude Code + Cursor)
Compliance with contributing guidelines
- ✅ Solves a real problem (35-pattern catalogue from real production-AI failure incidents)
- ✅ Well-documented (README + INSTALL + CHANGELOG + STATUS + per-component docs in
docs/)
- ✅ Includes examples (per-skill rendering examples in each
SKILL.md)
- ✅ Safe (PreToolUse hooks block dangerous operations; no destructive default behavior)
- ✅ Portable (Claude Code + Cursor variants; underlying 7 MCPs work cross-platform via MCP standard)
No network requests outside Anthropic API
The plugin's MCPs make local-only system inspections (filesystem, journals, version files, JSONL logs). No outbound HTTP except via the underlying LLM provider when explicitly invoked.
Auto-update
Plugin manifest does not auto-update. Users explicitly install via /plugin install aufgaard or claude --plugin-dir.