agent-sanitizer/QUICK_START.md

🚀 Quick Start - Agent Sanitizer

For Claude Code Users

Turn your Claude Code logs into a shareable dataset in 3 steps:

Step 1: Run the Tool

uvx agent-sanitizer

That's it! The tool will:

• Auto-detect your ~/.claude/projects directory
• Scan for security issues
• Show you what it found
• Ask for confirmation before cleaning

Step 2: Review the Results

You'll see a report like:

Security Audit Results
┌────────────┬───────┬────────────┐
│ Category   │ Found │ Severity   │
├────────────┼───────┼────────────┤
│ Credentials│ 12    │ 🔴 Critical│
│ Names      │ 8     │ 🟡 Medium  │
│ Emails     │ 5     │ 🟡 Medium  │
└────────────┴───────┴────────────┘

Step 3: Share Your Dataset

Option A: Save Locally

uvx agent-sanitizer --output ./my-dataset

Option B: Upload to HuggingFace

# Login first
huggingface-cli login

# Sanitize and upload
uvx agent-sanitizer --upload yourusername/my-coding-dataset

---

Advanced Usage

Dry Run First

See what would change without modifying files:

uvx agent-sanitizer --dry-run

Specify Directories

uvx agent-sanitizer \
    --input ~/.claude/projects \
    --output ./clean-dataset

Non-Interactive Mode

For scripts/automation:

uvx agent-sanitizer \
    --input ~/.claude/projects \
    --output ./clean-dataset \
    --no-interactive

---

What Gets Cleaned

✅ Removed

• API keys (OpenAI, Anthropic, DeepSeek, etc.)
• Passwords and tokens
• Email addresses
• Phone numbers
• Personal names
• Cryptocurrency private keys

✅ Preserved

• Code examples and snippets
• Test data (Hardhat accounts, test seeds)
• Tool usage patterns
• Thinking traces
• Error handling flows
• Technical usernames

---

Example Output

After running, you'll have:

clean-dataset/
├── splits/
│   ├── train.jsonl    # 80% of your data
│   ├── val.jsonl      # 10% of your data
│   └── test.jsonl     # 10% of your data
└── audit_report.json  # What was found and cleaned

---

Using Your Dataset

Load in Python

from datasets import load_dataset

# If uploaded to HuggingFace
dataset = load_dataset("yourusername/my-coding-dataset")

# Or load locally
import json

with open("clean-dataset/splits/train.jsonl") as f:
    data = [json.loads(line) for line in f]

print(f"Loaded {len(data)} coding interactions!")

Train a Model

Use the dataset to fine-tune your own coding assistant:

# Using transformers
python train.py \
    --dataset yourusername/my-coding-dataset \
    --model zenlm/zen-coder-14b \
    --output ./my-coding-assistant

---

Troubleshooting

"No agent logs found"

Make sure you're pointing to the right directory:

# For Claude Code
uvx agent-sanitizer --input ~/.claude/projects

# For Cursor
uvx agent-sanitizer --input ~/.cursor/logs

# For Continue
uvx agent-sanitizer --input ~/.continue/sessions

"huggingface_hub not found"

Install it:

pip install huggingface_hub

Or use uvx with dependencies:

uvx --with huggingface_hub agent-sanitizer --upload user/dataset

"Permission denied"

Login to HuggingFace first:

huggingface-cli login

---

Community

Share your dataset and help grow the community!

1. Tag it: Add agent-coding-dataset tag on HuggingFace
2. Link back: Mention Agent Sanitizer in your dataset card
3. Share: Post on Twitter/X with #AgentSanitizer

Example datasets:

• zenlm/agent-coding-dataset - 161k Claude Code interactions

---

Next Steps

• ⭐ Star the repo: https://github.com/zenlm/agent-sanitizer
• 📚 Read full docs: https://github.com/zenlm/agent-sanitizer#readme
• 🐛 Report issues: https://github.com/zenlm/agent-sanitizer/issues
• 💬 Discuss: https://github.com/zenlm/agent-sanitizer/discussions

Happy dataset sharing! 🎉