ExJtac/tokenstretch

Stop paying for wasted AI coding credits. One command finds the waste, explains it, and fixes it, with undo.
tokenstretch is free and open source (MIT). It runs 100% on your machine. There's no account, no sign-up, no telemetry, and no server. It looks at how Claude Code, Codex, and other coding agents are set up on your computer, shows you in plain English where your credits are going, and applies a small set of safe fixes. It backs up every file first, and tokenstretch undo puts everything back.
It doesn't replace your tools. It checks, configures, and measures them, including RTK and Caveman.
The local dashboard (tokenstretch dashboard), shown with sample data.
## What you get
A plain-English checkup. doctor runs 40+ local checks: the top model set as your everyday default, giant sessions that re-read hundreds of thousands of tokens on every step, duplicate instruction files you pay for twice, telemetry quietly left on, proxy chains, and setups that risk your provider's terms.
Fixes it applies for you. setup explains each change (what it is, why it costs you, which file and key it touches, and how to undo it), asks once, then applies it with live progress.
A live view. status shows what's applied, which model each tool defaults to, and how much history your newest sessions re-read on every step. It flags sessions over ~150K tokens with a "/clear at the next task break" tip.
A live watch. watch runs quietly in a terminal and warns you, once per problem, when a session's history gets too big, the top model is doing simple steps, the same file keeps getting re-read, or history is growing fast. Each warning says what's wrong, why it costs credits, and the one-line fix.
An instruction-file trimmer. trim lists every instruction file your agents resend on each turn (CLAUDE.md, AGENTS.md, GEMINI.md, copilot-instructions.md, and their imports) with approx. token counts, finds duplicated rules, oversized files, pasted logs and code, and stale paths, and proposes a trimmed version that only makes safe mechanical changes. It's a dry run unless you pass --apply, and undo restores it.
A model picker. model suggest " " recommends light, mid, or top for a task and names the model for each tool you have. model default --tier mid changes your default model, with a backup and undo.
A shareable report. report writes a single self-contained HTML file with the findings, the plan, what was applied, undo commands, and current status. It loads nothing from the internet.
A weekly report. report --weekly adds up your local usage logs per week: tokens by type and by model, your biggest session histories, waste spots, and a before/after view since setup. Cost appears only where your logs report it.
A local dashboard. dashboard opens one live page on your own machine with all of the above, updated every 5 seconds. See Dashboard.
Real measurements, never guesses. bench compares before and after on a throwaway copy of your repo, using your agent's own usage logs. If there's no data, it says no measurement.
Full undo. Every changed file is backed up under ~/.tokenstretch/backups/, and each change is recorded in a manifest.