Free Tool to Cut LLM Costs 61% — Auto-Scatter Hook

Hey. I built a tool that saved me $1,588.56 in one week. I'm releasing it for free. This article explains what's in the package, how it works, and how to get it.

Leave a comment and I'll drop the GitHub link. I want to make sure people read this and understand what they're getting before they install it — not because it's complicated, but because understanding it means you'll actually use it right.

What It Is

The auto-scatter hook is a local Python server (scatter_server.py) that intercepts every Claude Code prompt before it reaches the main model. It calls Claude Haiku to decompose your raw prompt into 6 structured frequency bands (sinc JSON), then injects that structure as system context. The main model gets a fully specified prompt. You get a first-try answer instead of a clarification loop.

What's in the Package

The Two Operating Modes

Haiku API mode — uses Claude Haiku for scatter. Cost: $0.002 per call. Saves $0.08 per call. Net: $0.078 saved per prompt. This is the mode most people will use. All you need is an Anthropic API key.

Local model mode — uses the fine-tuned Qwen2.5-7B GGUF (included in the package). Zero API cost. Requires an NVIDIA GPU (RTX 3090 or better recommended). Runs at 290 tok/s on an RTX 5090. Scatter quality slightly lower than Haiku for edge cases, but excellent for standard prompts. At zero marginal scatter cost, monthly savings project to $1,500+ for my workflow volume.

What to Expect After Installing

If your current exchange rate is near 4.2 (typical for unstructured Claude Code workflows), expect it to drop to 1.6-1.8 within the first day. The improvement is immediate — not gradual. Every prompt gets structured from the first hook call.

Total latency increase per prompt: 400-900ms (Haiku API call). This is front-loaded — you wait a bit more for the first response, but you wait much less overall because you don't have clarification loops. Net time per task decreases significantly.

The hook adds no observable memory overhead and doesn't modify any Claude Code files — it's a side-car pattern. Hook registration is in settings.json only. Uninstall is removing those 3 lines.

Prerequisites

# Requirements
Python 3.9+
Anthropic API key (for Haiku mode)
Claude Code (for hook integration)

# Install
pip install -r requirements.txt
uvicorn scatter_server:app --port 8461 &

# Add to Claude Code settings.json
# (see README for exact snippet)

Numbers from 7 Days of Production Use

21,194 prompts scattered. Exchange rate: 4.2 → 1.6. Cost saved: $1,588.56. Haiku overhead: $42.39. SNR improvement: 0.003 → 0.855. ROI: 38x. The hook has been running continuously since I deployed it — no crashes, no blocking incidents (the non-blocking fallback has never been needed).

Leave a comment and I'll send you the GitHub link. I'm releasing this free because I think it's genuinely useful and because the more people use it, the better the training data gets for the local model fine-tuning.