Show HN: LLM Anon – Anonymize code before sending to AI assistants

3 days ago 3

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

macOS/Linux:

curl -fsSL https://ollama.ai/install.sh | sh
Windows: Download from ollama.ai
Start Ollama service:
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

Quick Start (Recommended)

Install UV (if not already installed):

# macOS/Linux curl -LsSf https://astral.sh/uv/install.sh | sh
Install LLM Anonymizer:
Verify installation:

That's it! The llm-anon command is now available globally.

Alternative Installation Methods

Development Installation (click to expand)

For development or if you want to modify the code:

Clone this repository:

git clone https://github.com/ChristianBako/LLM-Anonymizer-.git cd LLM-Anonymizer-
Install dependencies:
Use with uv run:

uv run python -m llm_anon.cli --help

# Anonymize a single file and print to stdout llm-anon example.py # Anonymize and save to file llm-anon example.py -o anonymized.py # Process entire directory llm-anon src/ -r -o anonymized_output/

Development Usage (if installed from source)

# Use with uv run for development uv run python -m llm_anon.cli example.py

-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)

# Use different model with higher creativity llm-anon code.py -m codellama -t 0.3 # Preserve important strings and comments llm-anon api.py --preserve-strings --preserve-comments # Process entire project with verbose output llm-anon ./src -r -v -o ./anonymized # Use validation to ensure sensitive strings are removed llm-anon examples/lamasoft_example.py --validation-config examples/banned_strings.txt -v # Process with custom retry limit for stubborn validations llm-anon examples/lamasoft_example.py --validation-config examples/banned_strings.txt --max-retries 5

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

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). Comments start with #:

# Quick example echo "MyCompany" > banned_strings.txt echo "secret_api_key" >> banned_strings.txt echo "internal.company.com" >> banned_strings.txt # Or use the provided example cp examples/banned_strings.txt my_company_secrets.txt # Edit my_company_secrets.txt with your specific terms

Or create a comprehensive config file:

# company_secrets.txt # Company names and branding MyCompany mycompany.com # API keys and credentials API_KEY_12345 secret_password_2024 # Contact information [email protected] 1-800-MYCOMPANY # Product-specific terms MyCompanyCustomer MyCompanyLicenseManager

Pro tip: Start with your company name, domain, and any API keys or internal URLs. Check out examples/ for test files and sample configurations.

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

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

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

# Check if Ollama is running curl http://localhost:11434/api/tags # Restart Ollama service pkill ollama && ollama serve # List available models ollama list

# Pull the default model ollama pull llama3.2 # Or specify a different model llm-anon code.py -m llama3.1