@tomsmart-ai/mcp-audit-claude-code

MCP server for auditing Claude Code configurations. Originally built for AI Orchestration Audit-as-a-Service deliverables. Open-source under MIT.

Drop this server into your own Claude Code setup and it gains three tools for self-auditing your ~/.claude/settings.json + skills + hooks directories.

What it audits

audit_config_file

Parses your Claude Code settings.json and flags:

• HIGH: bypassPermissions defaultMode (security risk), hardcoded API keys leaked in config
• MED: missing permissions block, empty allow lists
• LOW: missing hooks block, no env vars set, excessive plugins enabled

audit_skills_dir

Scans ~/.claude/skills/ and flags per-skill:

• MED: missing SKILL.md, missing YAML frontmatter, missing description field
• LOW: description shorter than 20 chars, body shorter than 100 chars

audit_hooks_dir

Scans ~/.claude/hooks/ shell scripts and flags:

• HIGH: hook not executable (chmod +x missing), leaked API keys in hook content
• MED: missing set -e error handling in bash hooks
• LOW: hardcoded /Users/<name>/ paths (portability hit)

Install

npm install -g @tomsmart-ai/mcp-audit-claude-code

Use with Claude Code

Add to ~/.claude.json:

{
  "mcpServers": {
    "audit-claude-code": {
      "command": "mcp-audit-claude-code"
    }
  }
}

Restart Claude Code. Then ask:

Audit my Claude Code config at ~/.claude/settings.json

Claude calls audit_config_file, returns findings list with severity + fix hints.

Use case: AAA self-audit

This server powers the AI Orchestration Audit-as-a-Service stack audits (Tier 1 $500 — Tier 3 $1,500). When you commission an audit, this MCP server runs against your stack first, then a human review applies cultural context and shipping pragmatism on top.

If you want to skip the audit-as-a-service and run the same checks yourself, install this server and ask Claude to audit. The tools are the same; what you don't get is the cultural review layer + the implementation effort (audit-as-a-service includes one fix on Tier 2+).

Why open-source

The audit logic itself is straightforward — read JSON, check patterns, return findings. Selling closed access to the tool would protect zero moat. The moat is the cultural-context review + the experience auditing many stacks + knowing which findings actually matter for which workflows.

Open-source the tool, sell the judgment.

Roadmap

• v0.1.0 (today, Wt 26.05.2026) — initial audit tools for config, skills, hooks
• v0.2.0 — add audit_memory_dir for ~/.claude/projects/*/memory/ quality scan
• v0.3.0 — add audit_mcp_server_config for inspecting MCP server registrations
• v0.4.0 — add recommend_hooks returning suggested hook scaffolds for missing categories
• v0.5.0 — cross-check audit_config_file against audit_hooks_dir (consistency: hooks referenced in config exist on disk)

License

MIT

Author

Tom Smart — smartflowproai.com · @TomSmart_ai · github.com/smartflowproai-lang

Pair-built with Claude (Anthropic's coding agent). Tom scoped the tools, drafted findings categories, decided severity thresholds. Claude wrote the TypeScript. Tom reviewed each change and ran smoke tests before publish.