TokenTamer: The Proxy That Halves Your AI Costs (Without Cutting Corners)

If you've ever used Aider, Cursor, or Claude Code to edit your codebase, you know the problem: the agent reads the same file three times in one conversation, and you pay for it every time. TokenTamer sits between you and the API (literally — it's an HTTP proxy) and cleans house.

How it works is elegant: when the AI requests a file, TokenTamer parcels it up. It keeps function signatures, imports, class structure — everything the LLM needs to understand the code. Then it strips out function bodies you're not even touching. If the agent reads the same file again, TokenTamer knows what's stale and what's fresh, skeletonizing only the redundant reads.

It's not magic, it's AST parsing — understanding code structure and being smart about what to throw away. The AI still sees everything it needs, but the token count plummets.

Installation is embarrassingly simple: clone the repo, pip install -e ., change your tool's API base URL to http://127.0.0.1:8000/v1, and you're done. For Aider, Cursor, or anything using a custom API endpoint: no certs, no weird config. For Claude Code and Codex CLI (which hardcode the endpoint), there's a slightly more involved HTTPS setup, but still minutes.

Here's where it gets good: TokenTamer ships with a real-time dashboard showing exactly how many tokens you're saving and how much money stays in your pocket. Instant feedback matters — when you see "250 tokens saved" scrolling across the screen, you immediately know the tool is pulling its weight.

There are limits — it's still alpha, the savings numbers come from synthetic test payloads (not real long Claude Code sessions), and cross-request caching isn't here yet. But the roadmap is clear: Tree-sitter for proper multi-language AST, per-file heatmap dashboard, multi-turn caching. The dev knows where this is going.

Bluntly: if you use AI agents for coding and read the same file multiple times per session — spoiler, you do — this saves real money, right now.