My token bill says my development runs burned over 100 million tokens of inference. The budget layer in this system exists because I am the one paying that bill.
A 5-stage pipeline around Claude Code. Budget limits that can't be bypassed, quality gates that run before you see anything, and memory that grows between sessions.
Everything below was shipped solo by Michael Lip
The row below is my own ledger. The failures are printed in it too.
10 live products, all built with this pipeline
Most visitors land on one calculator here, use it, and leave. Good. The tools are free and stay free. I also run 40 Chrome extensions, and the machinery that keeps the whole portfolio going is for sale as three pipelines, each bought once, source included. I logged about 1,000 hours building the three of them, and my token bill says their development runs burned over 100 million tokens of inference. What you would buy is the machinery with its evidence shown, not a traffic outcome.
All three live at ml0x.com/x, with each refund window stated again on its own product page. If none of this is your problem, keep the calculators. They stay free either way.
The system is built for one person and sold once. The documents that ship with it specify every gate and every threshold. Read the free calculators on this site first if you want to judge the code before paying anything.
Without routing, every query hits Opus at $75/M output. With ml0x, 74% go to Flash at $0.28/M. You'll recoup the $99 in your first afternoon.
RouteLLM (ICLR 2025), Anthropic caching docs
The loop I run my own 40 extensions and 33 sites on, packaged and pinned at a version. The core is MIT and free on GitHub, and the product page says so before it asks for anything. No subscription. The calculators stay free.
