drykit

ESLint catches syntax errors. TypeScript catches type errors. Drykit catches "AI just created Modal2."

A linter for AI behavior. Pure Node CLI — zero AI calls, zero API keys, zero tokens spent on the tool itself.

Your AI assistant creates ConfirmModal.tsx because it doesn't know Modal already has a confirmation variant. Drykit gives it a 200-token memory file instead of 50,000 tokens of file-walking — and blocks the duplicate at commit time.

Try it now (existing project)

cd your-react-project
npx drykit scan

That's it. Drykit scans your project, builds a registry, and tells you what it found. No install, no config, no commitment.

Ready to go all-in? npx drykit init sets up the full pipeline.

Three layers of protection

┌─────────────────────────────────────────────────────┐
│  1. PREVENTION (before AI generates)                │
│  Claude Code skill + .cursorrules + AGENTS.md       │
│  → AI checks registry BEFORE writing new files      │
├─────────────────────────────────────────────────────┤
│  2. DETECTION (at commit time)                      │
│  pre-commit hook → drykit check --ci                │
│  → blocks duplicates, unregistered files, secrets   │
├─────────────────────────────────────────────────────┤
│  3. VISIBILITY (ongoing)                            │
│  drykit health / drykit stats                       │
│  → god-objects, missing docs, token savings counter │
└─────────────────────────────────────────────────────┘

Duplicates — catches ConfirmModal when Modal variant="confirmation" exists. Unregistered files — catches components that slipped past the registry. Secrets — catches hardcoded Stripe keys, AWS tokens, passwords in source files.

The token math

Without drykit — AI orients itself every session:

| Action | Tokens | |---|---| | Read 10–20 files to figure out what exists | 30k–80k | | Re-read the same files next session | 30k–80k | | Per week (~20 sessions) | ~1M tokens |

With drykit — AI reads one 200-token fingerprint:

| Action | Tokens | |---|---| | Read .drykit/fingerprint.md | ~200 | | Read front.md or api.md on demand | ~1–3k | | Per week (~20 sessions) | ~20–60k tokens |

~95% fewer tokens on orientation (estimate). The real win is accuracy — fewer wrong turns, fewer duplicates, faster sessions.

Before vs. after

Before:

src/components/
├── Modal.tsx
├── Modal2.tsx               ← "quick tweak"
├── ConfirmModal.tsx         ← AI didn't know Modal existed
├── DeleteModal.tsx          ← another AI session
├── ConfirmDeleteModal.tsx   ← AI made this last Tuesday
└── Dialog.tsx               ← somebody googled "React dialog"

After:

src/components/
└── Modal.tsx                ← variants: primary | confirmation | form | delete
                                registered, documented, AI knows it exists

Commands

| Command | What it does | |---|---| | drykit init | Detect structure, configure all AI tools, install hooks | | drykit scan | Rebuild registry + fingerprint from source | | drykit add <Name> | Scaffold + register a new component | | drykit check | Duplicates + unregistered + secrets scan | | drykit check --json | Structured JSON output for CI/tooling | | drykit check --ci | Exit codes: 0=clean, 1=warnings, 2=errors | | drykit check --report | Markdown report to stdout | | drykit check --report-file <path> | Markdown report to file (for PR comments) | | drykit merge <A> <B> | Merge duplicate components into one | | drykit health | Large files, missing docs, progress counter | | drykit stats | Cumulative token savings (estimate) | | drykit docs | Generate COMPONENTS.md from registry | | drykit eject | Remove all drykit files |

What gets generated

your-project/
├── drykit.config.mjs              # scan paths, DRY-risk keywords
├── src/registry.json              # source of truth
├── .drykit/
│   ├── fingerprint.md             # ~200 tokens — AI reads this first
│   ├── front.md                   # UI inventory (on demand)
│   ├── api.md                     # routes + schemas (on demand)
│   └── savings.json               # cumulative token savings
├── .cursorrules                   # Cursor reads this
├── AGENTS.md                      # Codex / Kiro reads this
├── CLAUDE.md                      # drykit section appended
├── .claude/
│   ├── skills/drykit/SKILL.md     # proactive registry check
│   ├── agents/drykit-scanner.md   # Haiku subagent
│   ├── agents/drykit-architect.md # Sonnet subagent
│   └── hooks/drykit-pretooluse.mjs # blocks writes before they happen
├── .kiro/
│   ├── steering/drykit.md         # always-on rules
│   └── steering/drykit-front.md   # auto-updated on scan
└── .husky/pre-commit              # drykit check --ci

How it works

1. drykit scan walks your source, extracts names/props/variants/dependencies, writes registry.json
2. .drykit/fingerprint.md is a 200-token summary the AI reads at session start — instead of reading 20 files
3. AI tools (Claude Code, Kiro, Cursor) get rules injected: "check the registry BEFORE creating"
4. drykit check --ci runs at pre-commit — blocks duplicates, unregistered files, and hardcoded secrets
5. drykit check --report generates a markdown report — duplicates, inconsistent prop usage, secrets, unregistered
6. drykit stats tracks how many issues were caught over time, with weekly deltas

The entire pipeline is pure Node.js. No AI calls. No API keys. No tokens spent on the tool itself.

Who this is for

For developers who ship with AI.

• Teams using Claude Code, Kiro, Cursor, or Copilot
• Projects with 10+ components where the AI keeps reinventing the wheel
• Anyone who's paid for an AI session that ended with "oh, we already had that"

Configuration

drykit.config.mjs — auto-detected, always editable:

export default {
  projectName: 'my-app',
  stack: 'Next.js 15 + Tailwind 4',
  scan: {
    components: ['src/components/**/*.tsx'],
    hooks:      ['src/hooks/**/*.ts'],
    utils:      ['src/utils/**/*.ts'],
    routes:     ['src/app/api/**/*.ts'],
    schemas:    ['src/schemas/**/*.ts'],
  },
  registry: 'src/registry.json',
  dryRisk: ['Modal', 'Form', 'Card', 'Button', 'Dialog', 'Drawer',
            'Toast', 'Dropdown', 'Select', 'Input', 'Table'],
};

FAQ

Does it slow down commits? No. Pure Node scan — milliseconds on typical projects.

React Native / Vue / Svelte? v1 is React-only. Other frameworks on the roadmap.

Does it replace my design system? No — it makes it impossible for AI to quietly ignore the one you already have.

Will it overwrite my CLAUDE.md? No. Appends a marked section, leaves the rest alone.

What if I don't use Claude Code? Registry + pre-commit + .cursorrules still work. The Claude-specific files are optional.

Requirements

• Node.js 20+
• A React project (Next.js, Vite, Remix — anything)

License

MIT