Audit your codebase from Claude Code. No API key.

Why use Claude Code as the runner

If you already pay for Claude.ai, your Pro / Max subscription unlocks a generous Claude Code budget. Anatoly runs on top of that runtime: every read, grep, file walk, RAG query and finding-resolution round-trip flows through Claude Code instead of an API key. Same audit pipeline. Zero tokens billed on top.

The first run is heavy — full embedding, all seven axes, deliberation pass. Every run after that is incremental: SHA-256 cached, unchanged files cost $0, and the hook only re-audits what you just edited.

One command to install the hook

$ npx anatoly hook init
→ writing .claude/settings.json (PostToolUse, Stop)
→ default anti-loop: max_stop_iterations 3
✓ hook installed · close & reopen Claude Code to activate

$ npx anatoly run
→ provider · claude-code (subscription)
→ axes (parallel) · sonnet 4.6 · 7/7 OK
→ deliberation · opus 4.7 · 22 → 11 confirmed
✓ .anatoly/runs/2026-05-05_104210/public_report.md

The save → audit → fix loop

Once the hook is installed, the loop runs automatically as you save files inside Claude Code. The anti-loop guard caps it at three cycles per file so it never spins on itself.

What the hook config looks like

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{ "type": "command",
        "command": "npx anatoly hook on-edit", "async": true }]
    }],
    "Stop": [{
      "hooks": [{ "type": "command",
        "command": "npx anatoly hook on-stop", "timeout": 180 }]
    }]
  }
}

What the integration gives you

• Hook-driven loop - install once, audits run on every Edit/Write. Anti-loop guard prevents runaway.
• Findings in your editor session - the agent surfaces newly-detected issues directly inside Claude Code, with file + line + evidence so you can fix or dismiss in place.
• Opus deliberation, on-subscription - the three-tier finding refinement pass (auto-resolution, inter-axis contradiction detection, then Opus with full tool access) runs on the same subscription budget.
• Same incremental cache - SHA-256 cached re-runs mean unchanged files cost nothing, regardless of where the inference lands.

Mix it with other providers

You're not locked in. Set the cross-cutting routing once, then override the heavier axes individually - useful when you want Opus on Correction but Sonnet everywhere else, or when you bring in a local model alongside Claude.

routing:
  generation:
    quality: anthropic/claude-sonnet-4-6
    deliberation: anthropic/claude-opus-4-6
    fast: anthropic/claude-haiku-4-5-20251001

evaluation:
  axes:
    correction:
      enabled: true
      model: anthropic/claude-opus-4-6    # override quality routing
    utility: true    # inherits routing.generation.quality

Who this fits

• Senior engineers — maintaining a brownfield repo, who want a continuous audit loop without spinning up a separate API budget.
• Solo founders — shipping fast with Claude Code, who want to know what their codebase actually contains.
• Auditors & consultants — billing fixed-fee engagements, who'd rather amortize a Claude Max subscription than a per-token bill.