Refactory

Hybrid code decomposition. AI plans the boundaries. A deterministic engine handles the routine extractions. Minimize tokens, maximize syntax validity.

Refactory splits monolith source files into clean modules. It uses an LLM for one thing — deciding which functions group together. Everything else is mechanical: function boundary detection, import resolution, module assembly, syntax validation, scoring.

JavaScript and Python extraction is mostly mechanical. The deterministic engine handles the straightforward moves — the routine 80% that's a waste of AI time and tokens. The LLM still handles complex edge cases where judgment matters. Other languages use LLM extraction with adaptive compression.

Works with Claude Code, Cursor, Windsurf, VS Code Copilot — any MCP client. Or use the CLI directly.

Results

Tested against 15 production monoliths:

Related MCP server: sourcebook

Quick Start

MCP (recommended)

Add to your .mcp.json:

{
  "mcpServers": {
    "refactory": {
      "command": "npx",
      "args": ["@refactory/mcp"],
      "env": {
        "GROQ_API_KEY": "your-key-here"
      }
    }
  }
}

Then tell your AI tool: "Analyze and decompose src/big-file.js into modules"

One free API key (Groq or Gemini) is needed for the PLAN step only. Extraction is mechanical — no key required for JS/Python.

CLI

git clone https://github.com/codedrop-codes/refactory.git
cd refactory && npm install
node src/cli.js decompose src/big-file.js

Other commands:

refactory analyze src/big-file.js        # Health check + function map
refactory plan src/big-file.js           # Generate module boundaries (needs LLM key)
refactory verify lib/modules/            # Check extracted modules
refactory languages                      # Show supported languages
refactory providers                      # Show configured LLM providers
refactory test submit broken.js          # Submit a file that breaks extraction
refactory test run                       # Validate preprocessors against test corpus

How It Works

  1. ANALYZE         Scan functions, dependencies, health — mechanical
       |
  2. CHARACTERIZE    Snapshot exports before touching anything — mechanical
       |
  3. PLAN            LLM decides module boundaries — the only AI step
       |
  4. EXTRACT         Copy functions by line range, resolve imports — mechanical
       |               (LLM fallback for unsupported languages)
  5. FIX-IMPORTS     Rewrite require()/import paths — mechanical
       |
  6. VERIFY          Syntax check, load check, export comparison — mechanical
       |
  7. METRICS         Refactory Score + HTML report — mechanical

6 of 7 steps are deterministic. The LLM only decides where to split — it never touches your code.

Language Support

Mechanical extraction handles the routine cases: the preprocessor finds function boundaries by parsing, copies them by line range, and resolves imports deterministically. Complex patterns (dynamic exports, deeply interleaved logic) still go through the LLM.

Contribute a preprocessor for your language.

Refactory Score

A single number (0.0 to 1.0) that measures decomposition quality.

Score = clean_rate × size_reduction

• clean_rate — modules that load without errors / total modules

• size_reduction — 1 − (largest module / original file)

A score of 1.0 means every module loads cleanly and no module is bigger than the original.

Provider Routing

You only need one free key for the PLAN step. Extraction is mechanical for supported languages.

Set at least one: GROQ_API_KEY, GOOGLE_API_KEY, OPENROUTER_API_KEY, or SAMBANOVA_API_KEY.

Test Corpus

Found a file that breaks extraction? Submit it:

refactory test submit broken-file.js -d "what went wrong"

Secrets are stripped automatically. Every submission becomes a permanent test case. The extractor gets stronger with every report.

Report via GitHub if you prefer.

Community

• Discord — Help, ideas, show your results

• Discussions — Feature requests, language requests

• Issues — Bug reports

• Contributing — Build a preprocessor, submit test files

License

AGPL-3.0 — see LICENSE.

Premium language packs available under commercial license. See refactory.codedrop.codes.