Anatoly

"Can I clean here?"

An open source, no-frills code reviewer that out-lifts the giants without raising its voice.

Anatoly is a deep-audit agent for existing codebases. It evaluates every file on seven axes: silent bugs, dead code, semantic duplicates, over-engineering, weak tests, broken best practices, and stale documentation. Every finding is proven with a file, a line, and the evidence trail that led to it. No code is modified.

Built for the codebases that grew faster than anyone could review them, especially the ones shipped with Cursor, Claude Code, Lovable, v0, Bolt, or Windsurf.

npx @r-via/anatoly@latest run   # one command, full codebase audit

---

The problem

You shipped fast. Maybe with AI, maybe with a sprint that never ended. Now the codebase has:

• functions that do the same thing under different names,
• exports nobody imports,
• silent logic bugs the type system can't catch,
• bespoke abstractions wrapping three values,
• "tested" code with no behavioral coverage,
• best-practice rules the team agreed on but stopped enforcing,
• public APIs whose /docs/ pages drifted out of sync.

Linters won't find this. Type checkers won't either. A senior reviewer would, if they had a week. Anatoly does it in minutes, and gives you a report you can defend.

---

What a run looks like

You run one command. A few minutes later, you get a report like this:

15 files reviewed in 10 min, $6.13 in AI analysis so you don't have to. Verdict: NEEDS_REFACTOR · 26 findings in 11 files

Axis	Health
Correction	🟥🟥🟥🟥🟥🟥🟥🟥🟥⬜ 93% OK
Utility	🟥🟥🟥🟥🟥🟥🟥🟥⬜⬜ 83% used
Duplication	🟥🟥🟥🟥🟥🟥🟥🟥🟥⬜ 90% unique
Overengineering	🟩🟩🟩🟩🟩🟩🟩🟩🟩⬜ 90% lean
Best Practices	🟥🟥🟥🟥🟥🟥🟥⬜⬜⬜ avg 7.4 / 10

And findings like these, each one with file+line, a verdict, and the evidence that produced it:

• 🐛 src/engine.ts computePayout: the house-edge multiplier sign is inverted: it boosts payout instead of reducing it. RTP is broken.
• 🐛 src/rng.ts weightedPick: Math.random() used as the RNG source in a function whose own docstring says "suitable for gaming RNG applications".
• ♻️ src/types.ts LegacySpinResult: exported, imported by 0 files.
• 📋 src/engine.ts checkLine ⇄ src/paytable.ts lineWins: semantically identical (RAG cosine 0.852): same WILD-skip, same consecutive-match loop, different names. A textual diff would never see it.
• 🏗️ src/engine.ts EngineContainer: bespoke IoC container backed by a stringly-typed Map<string, unknown>, holding three values.

One command. Full codebase. Evidence-backed findings. No code modified.

Languages: TypeScript, Python, Rust, Go, Java, C#, Bash, SQL, YAML, JSON.

---

Who runs Anatoly

If you are…	You use Anatoly to…
A senior dev or tech lead shipping AI-generated code	Surface the duplicates, dead code, and silent bugs your AI tools left behind.
A team on a Claude or Cursor subscription	Get Opus-grade deliberation on every audit, at $0 per token.
A regulated org (finance, health, defense)	Audit air-gapped: embeddings local, LLM-optional local, no source byte leaves your network.
An M&A or legal expert doing tech due diligence	Produce reports that survive counter-expertise: every claim pinned to file, line, and evidence.
A DevOps team running CI/CD quality gates	Audit every PR with structured JSON output for your pipeline.

---

Why the findings hold up

• Evidence-backed: every finding is grep'd, read, and cross-referenced before it's reported. No hallucinated issues.
• Semantic duplicate detection: finds functions that do the same thing even when they look nothing alike.
• Deliberation pass: multi-LLM arbitration on disputed findings.
• $0 on unchanged code: SHA-256 cache. Re-running on a clean repo costs nothing.
• Run and walk away: Telegram when done. Watch mode and Claude Code hook for write → audit → fix loops.
• Local by default: embeddings run on your machine. With Claude Code subscription, no API key needed.

Fair warning: this is context-and-time-hungry by design. Dream-quality code doesn't come cheap. It takes serious effort (and tokens) to get there.

---

Measured accuracy

Anatoly is benchmarked against anatoly-bench: fixtures seeded with curated defects (bugs, dead code, semantic duplicates, over-engineering) and scored by F1 against a ground-truth catalog.

xychart-beta
    title "slot-engine fixture, global F1 (%) per bench run"
    x-axis [v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v37b, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v47b, v48, v49, v50, v50b, v51, v52, v52b, v53, v54, v56, v57, v58, v59, v60, v61, v62, v63, v64, v65, v66, v67]
    y-axis "F1 (%)" 0 --> 100
    line [40.7, 43.2, 46.8, 43.5, 56.8, 65.5, 62.7, 61.0, 65.0, 57.8, 67.8, 63.8, 72.9, 64.2, 69.9, 65.5, 71.4, 69.7, 67.0, 66.2, 69.4, 68.8, 68.7, 67.8, 58.3, 56.2, 67.3, 71.5, 57.9, 62.1, 71.6, 57.5, 71.5, 70.4, 74.3, 65.8, 70.2, 72.2, 67.9, 67.6, 68.0, 76.5, 72.3, 62.8, 74.6, 75.5, 70.1, 75.5, 62.8, 70.9, 69.7, 75.8, 78.3, 73.2, 79.5, 76.3, 73.5, 76.8, 61.4, 73.2, 70.9, 69.5, 66.0, 69.9, 66.9, 72.9, 69.9, 75.1, 68.6]

Loading

The fixture (27 seeded defects) has been frozen since v16, so every later point is Anatoly behaviour plus run-to-run LLM variance (about 9 points on a single run). Best run: 79.5% (v52b). The last twelve points are the move to the Claude 5.5 generation shipped in 0.11.0: at equal reasoning effort and after the prompt recalibration, Sonnet 5.5 scores 69.9% and 75.1% (v65, v66), on par with the Sonnet 4.6 controls of the same day (72.3% mean) at about 40% of the cost and half the wall time.

Per-axis curves, change-log, and remaining misses: slot-engine results.

---

Getting started

# One-off run
npx @r-via/anatoly@latest run

# Or install globally
npm install -g @r-via/anatoly
anatoly run

Prerequisites: Node.js >= 20.19, plus one LLM auth path:

• Claude Code subscription (default): no API key needed
• BYOK: ANTHROPIC_API_KEY, etc.
• Local: Ollama, LM Studio, vLLM

Default models (since 0.11.0): Claude Sonnet 5.5 for the review axes, Haiku 5.5 for the mechanical ones, Opus 5.5 for deliberation, every call at high reasoning effort. Models and effort are set per tier in .anatoly.yml.

See Installation and Recommended LLM Setup.

Optional: GPU embeddings

For better duplication detection, set up GGUF mode (Docker + NVIDIA GPU ≥ 12 GB VRAM):

npx anatoly local-embeddings upgrade

---

Common commands

npx anatoly run                # Full audit
npx anatoly watch              # Daemon: incremental re-audit on change
npx anatoly estimate           # Forecast tokens, cost, time (no API calls)
npx anatoly runs attach        # Inspect a run (live or finished)
npx anatoly reset              # Wipe all state

Full CLI: docs/03-CLI-Reference/. Configuration: docs/01-Getting-Started/02-Configuration.md.

---

Documentation

Section	Highlights
Getting Started	Vision, Installation, Configuration
Architecture	Pipeline, 7-Axis System, RAG Engine, Deliberation
CLI Reference	Commands, Global Options
Integration	Claude Code Hooks, CI/CD, Watch Mode, Telegram
Design Decisions	Why Local RAG, Evidence-Based

---

Development

git clone https://github.com/r-via/anatoly.git
cd anatoly
npm install
npm run dev        # Run with tsx
npm run build      # Build with tsup
npm test           # Vitest
npm run typecheck  # tsc --noEmit

Contributing

1. Fork, branch (feat/my-feature)
2. npm test && npm run typecheck
3. Open a PR

By contributing, you agree to the Contributor License Agreement (Anatoly uses dual licensing: AGPL-3.0 + commercial).

License

Dual-licensed:

• Open source: AGPL-3.0. Free for open source, personal, and academic use.
• Commercial: proprietary license available. See COMMERCIAL.md.

Contact

• Website: anatoly.cloud
• Issues: github.com/r-via/anatoly/issues