llm-anon 0.1.0

CLI tool to anonymize code using local LLM before sending to Claude Code

LLM Anonymizer

A CLI tool to anonymize code using a local LLM before sending to Claude Code or other AI services.

Prerequisites

Install Ollama

1. macOS/Linux:

   curl -fsSL https://ollama.ai/install.sh | sh
   

2. Windows: Download from ollama.ai

3. Start Ollama service:

   ollama serve
   

4. Install a model (in a new terminal):

   # Install Llama 3.2 (recommended)
   ollama pull llama3.2
   
   # Or install other models
   ollama pull codellama
   ollama pull llama3.1
   

Install UV (if not already installed)

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Installation

1. Clone this repository:

   git clone <repository-url>
   cd llm-anon
   

2. Install dependencies:

   uv sync
   

Usage

Basic Usage

# Anonymize a single file and print to stdout
uv run python -m llm_anon.cli example.py

# Anonymize and save to file
uv run python -m llm_anon.cli example.py -o anonymized.py

# Process entire directory
uv run python -m llm_anon.cli src/ -r -o anonymized_output/

Options

• -o, --output PATH: Output file or directory
• -m, --model TEXT: LLM model to use (default: llama3.2)
• -t, --temperature FLOAT: Temperature for generation (default: 0.1)
• --preserve-comments: Keep original comments
• --preserve-strings: Keep string literals unchanged
• -r, --recursive: Process directories recursively
• -v, --verbose: Show detailed progress
• --validation-config PATH: Path to validation config file with banned strings
• --max-retries INTEGER: Maximum retries for validation failures (default: 3)

Examples

# Use different model with higher creativity
uv run python -m llm_anon.cli code.py -m codellama -t 0.3

# Preserve important strings and comments
uv run python -m llm_anon.cli api.py --preserve-strings --preserve-comments

# Process entire project with verbose output
uv run python -m llm_anon.cli ./src -r -v -o ./anonymized

# Use validation to ensure sensitive strings are removed
uv run python -m llm_anon.cli code.py --validation-config banned_strings.txt -v

# Process with custom retry limit for stubborn validations
uv run python -m llm_anon.cli sensitive_code.py --validation-config company_secrets.txt --max-retries 5

Supported Languages

• Python (.py)
• JavaScript (.js, .jsx)
• TypeScript (.ts, .tsx)
• Java (.java)
• C/C++ (.c, .cpp, .cc, .cxx, .h, .hpp)
• Rust (.rs)
• Go (.go)

Validation System

The tool includes a powerful validation system to ensure sensitive information is completely removed from anonymized code.

Creating a Validation Config

Create a text file with banned strings (one per line):

# company_secrets.txt
# Company names and branding
MyCompany
mycompany.com

# API keys and credentials  
API_KEY_12345
secret_password_2024

# Contact information
[e-mail]
1-800-MYCOMPANY

# Product-specific terms
MyCompanyCustomer
MyCompanyLicenseManager

How Validation Works

1. Initial Anonymization: LLM processes code normally
2. String Detection: Scans output for banned strings using word boundaries
3. Re-prompting: If banned strings found, sends explicit removal instructions
4. Retry Logic: Repeats up to --max-retries times until validation passes
5. Failure Handling: Reports specific banned strings if validation ultimately fails

Validation Features

• Case-sensitive matching by default
• Word boundary detection prevents false positives
• Progressive prompting gets more explicit with each retry
• Detailed error reporting shows exactly which strings were found

How It Works

1. File Detection: Automatically detects programming language from file extension
2. LLM Processing: Sends code to local Ollama model with anonymization prompt
3. Validation (if enabled): Checks output against banned strings and re-prompts if needed
4. Smart Replacement: Replaces variable names, function names, and identifiers while preserving:
   • Code structure and logic
   • Control flow
   • Data types
   • Import statements
   • Syntax and formatting

Troubleshooting

Ollama Connection Issues

# Check if Ollama is running
curl http://localhost:11434/api/tags

# Restart Ollama service
pkill ollama && ollama serve

# List available models
ollama list

Model Not Found

# Pull the default model
ollama pull llama3.2

# Or specify a different model
uv run python -m llm_anon.cli code.py -m llama3.1

Performance Tips

• Use smaller models for faster processing: llama3.2:1b
• Lower temperature (0.1) for more consistent results
• Process files individually for large codebases to avoid timeouts