epoko77-ai

epoko77-ai / tokensave

Public

멀티에이전트 Claude Code 하네스를 위한 토큰 비용 lint — 10 카테고리 × 31 sub-patterns 분류학 + 정적 감사 스크립트 + LLM 0회 결정 트리. Korean primary, English mirror.

github.comepoko77-aitokensave agentic-ai ai-cost ai-engineering anthropic claude-code
16
2
85% credibility
Found May 17, 2026 at 17 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

tokensave is a free tool that helps people who use Claude Code save money on their AI costs. It works like a financial auditor for your AI agent setup: first, it scans all your agents to find patterns that waste money (like using expensive AI models for simple tasks that cheaper models could handle). Second, it provides a decision guide that tells you which AI model to use for any given task. The tool runs entirely on your own computer using only Python's built-in features, so no external services or payments are needed. It includes detailed reference materials about cost optimization patterns, all backed by official documentation and community research.

How It Works

1
💸 You notice your AI bill is getting too high

You've been using Claude Code with multiple AI agents, and the monthly costs have become a concern.

2
🔍 You run a quick check on your setup

You run a simple scan that looks through all your AI agent configurations to find where money might be slipping away.

3
😱 The report shows expensive mistakes

The results reveal that almost all your agents are using the most expensive AI model for simple tasks, and none of them are using cost-saving tricks.

4
You choose how to fix it
🔧
Follow the guided fixes

The tool shows you exactly which agents to change and what to switch them to, with clear explanations.

📋
Look up the right model yourself

You describe a task in plain language and get an instant recommendation for which model fits best.

5
✨ You apply the changes

You update your agent configurations based on the recommendations, switching some to cheaper models.

🎉 Your costs drop significantly

The same work gets done at a fraction of the cost, and you have a reference guide for future projects.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 17 to 16 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is tokensave?

tokensave is a cost-optimization linter for Claude Code multi-agent setups. It scans your agent configurations, skill files, and harness definitions to catch the five patterns that make multi-agent teams expensive but unjustified. Run `audit.py` on your home directory and get a report in seconds, zero LLM calls required. The tool ships a decision tree you can query from the command line to ask "what model fits this task?" and get Python/Haiku/Sonnet/Opus recommendations backed by a 24-row task-to-model matrix. Everything is standard library Python with no pip install.

Why is it gaining traction?

The hook is stark numbers: one operator's 27-harness catalog showed 99% of agents running Opus with zero harnesses using prompt caching, paying a 7x-15x multiplier without applying the patterns that justify it. The taxonomy of 31 cost-burning sub-patterns across 10 categories gives you a shared vocabulary for what "well-formed" multi-agent design actually means. Most tools focus on individual file quality. This one measures the system.

Who should use this?

Claude Code operators running multi-agent teams who want to audit cost hygiene. If you have 5+ agents and have never thought about model tier mixing or prompt caching, this will surface the gaps. Teams optimizing existing harnesses will get the most value from the audit and hotspot prioritization. Single-user hobby setups with a few agents probably do not need this yet.

Verdict

The credibility score of 0.85% reflects a tool built from one operator's catalog with community validation still incoming. At 16 stars and v1.0, it is usable but early-stage. The reference documents alone are worth reading even if you never run the CLI. Jump in, submit your baseline, and help harden the taxonomy.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.