See inside your Claude Code. Fix what's broken. Share what works.

npx claude-code-xray

Claude Code has 70+ settings, 25 hook events, a 4-level instruction hierarchy, and a full permission system. Most setups use less than 5% of this surface.

Real consequences of bad setups:

• Claude Code deleted a user's entire Mac home directory (Dec 2025)
• terraform destroy wiped a production database (Mar 2026)
• CVE-2025-59536: RCE via malicious hooks in untrusted repos
• Claude co-authored commits leak secrets at 3.2% rate (2.4x human baseline)

X-Ray scans your setup, shows what's dangerous, and fixes it in one command.

# 1. Scan — see your score and what's wrong
npx claude-code-xray

# 2. Fix — apply safe, conservative fixes with backup + rollback
npx claude-code-xray fix --apply

# 3. Badge — show your score in your README
npx claude-code-xray badge

That's it. Three commands. No config, no signup, no data leaves your machine.

BEFORE                              AFTER (4 minutes later)

YOUR SCORE: 49/100                  YOUR SCORE: 73/100

Safety & Security  50/100 [!]       Safety & Security  100/100
Capability         25/100           Capability          50/100
Automation         75/100           Automation          75/100
Efficiency         79/100           Efficiency          79/100

The fix command adds deny rules for secrets, enables sandbox isolation, installs a PreToolUse safety hook, and patches the Bash deny gap. Dry-run by default. Every change shown as a diff with a "why this is safe" explanation.

Claude Code X-Ray ──────────────────────────────────────────

  YOUR SCORE: 49/100  (4/4 dimensions scored)

  Safety & Security    █████░░░░░   50/100  [!]
  Capability           ███░░░░░░░   25/100
  Automation           ████████░░   75/100
  Efficiency           ████████░░   79/100

┌─ WHAT YOU HAVE ────────────────────────────────────────────
│  ✓ Permission mode: default
│  ✓ PreToolUse safety hook: yes
│  ✓ MCP server trust: per-server
│  ✓ Cache hit ratio: 99%
│
├─ WHAT YOU'RE MISSING ──────────────────────────────────────
│  [!] No deny rules for .env, secrets, credentials
│  [!] sandbox.enabled is false (Bash bypasses deny rules)
│  [ ] Coordinator Mode available but not configured
│  [ ] 6 hook events uncovered
│
├─ WHAT TO DO NEXT (ranked by impact) ──────────────────────
│  +15-36 pts  Fix critical safety gaps    xray fix
│  +12-32 pts  Fix remaining gaps          xray fix
│
│  Fix all: npx claude-code-xray fix
└────────────────────────────────────────────────────────────

X-Ray is not a linter. It doesn't check your code style or flag unused imports. It checks your Claude Code harness — the configuration that controls what an AI agent can do on your machine.

Every check is labeled [VERIFIED] (from official schema/docs) or [INFERRED] (from source analysis), so you know the confidence level.

Skipped dimensions (no data) are excluded from the score. Weights renormalize automatically.

npx claude-code-xray fix           # dry-run: show what would change
npx claude-code-xray fix --apply   # apply fixes with backup + rollback

Fixes are conservative: dry-run by default, each change shown as a diff with a "why this is safe" explanation, automatic backup before applying, rollback on failure.

npx claude-code-xray badge         # markdown for README
npx claude-code-xray badge --svg   # standalone SVG

Add to your README:

npx claude-code-xray history       # score over time

Claude Code's source revealed autonomous background agents, multi-agent orchestration, and cloud compute are coming. When these ship, your setup needs to be safe. A background agent with bypassPermissions can modify any file at 3am.

X-Ray checks your readiness today. Fixes the gaps. Tracks your progress.

X-Ray reads your Claude Code configuration files:

• ~/.claude/settings.json (user settings)
• .claude/settings.json (project settings)
• .claude/settings.local.json (local overrides)
• ~/.claude.json (MCP servers, global config)
• ~/.claude/projects/*/ (session transcripts, usage only, never content)
• CLAUDE.md files (all 4 hierarchy levels)
• ~/.claude/skills/ (installed skills)

Privacy: Session transcript analysis ONLY reads message.usage fields (token counts). Message content is never read, stored, or transmitted. X-Ray runs entirely locally. No data leaves your machine.

• Node.js 18+
• Claude Code CLI installed
• Optional: gh CLI (for some capability checks)

Issues and PRs welcome. If you find a check that gives bad advice or a score that feels wrong, open an issue — X-Ray improves by community signal.

MIT