TokenTamer: The proxy that slashes your AI bills without breaking anything

Picture this: your AI coding agent (Aider, Cursor, Claude Code, whatever) fires off a 100k token request to help build a function. Half those tokens are background files the AI doesn't actually care about—library headers, helper function implementations, stuff the model already knows or doesn't need.

TokenTamer wedges itself between your agent and the LLM API like an intelligent gatekeeper. It intercepts the raw payload, parses the AST of your code files (meaning it understands the structure), and compresses "background" files into lean skeletons: keeps function signatures, imports, class structures, ditches the implementations. Boom. 15k tokens instead of 100k.

Setup? Trivial. It's a drop-in: change your agent's API base URL, point it at http://127.0.0.1:8000 and go. Aider: aider --openai-api-base http://127.0.0.1:8000/v1. For stubborn tools (Claude Code, Codex) that don't expose the parameter, TokenTamer also supports SSL interception—some certificate wrangling required, but nothing crazy.

The kicker: while the model chews on fewer tokens, you watch a real-time dashboard telling you exactly how much you're saving. We're talking $2.45 per request less. Multiply that by 100 requests a day and... yeah, you get it.

Supports Python natively (real AST parsing), JavaScript, TypeScript, Go, Rust, Java, C#—everything you'd actually use. Optional configuration via YAML if you want to fine-tune what gets preserved (docstrings, class attributes, etc.). Zero latency overhead: compression happens locally in milliseconds.

The catch? It's still young (5 stars on GitHub, but we're talking a side project), so expect to hit rough edges on weird codebases. Not a miracle-worker for absolute chaos, but for most real-world projects it's a solid win.