🚀 I just open-sourced something I built out of frustration — and it's saving me real money every day.https://github.com/borhen68/TokenTamer
If you use AI coding agents (Cursor, Claude Code, Aider…), you've probably noticed how fast your API bills stack up. The problem? Your agent keeps re-sending the same files — full source code — every single turn. You're paying for tokens you already paid for.
So I built TokenTamer — a drop-in proxy that sits between your agent and the LLM API and quietly cuts token usage by 50–80%. No config changes. No code changes. Just point your API base URL at it and go. ✅
Here's what it does under the hood:
🧠 AST-based compression — strips function bodies from background files, keeps only signatures. The LLM knows what exists, without reading every line.
🔧 Tool-aware compression — skeletonizes stale file reads, keeps the latest one intact.
💾 Prompt cache hijacking — injects Anthropic cache breakpoints so long Claude Code sessions hit the cache instead of paying full price (~73% off on long runs).
💰 Real-time dashboard — watch tokens saved and dollars saved live in your terminal.
It's MIT licensed, works with OpenAI + Anthropic APIs, and takes 5 minutes to set up.
👇 Star it, try it, break it — and tell me what you think.
Top comments (0)