tokenshame 0.1.0

Shame your AI agent for wasting tokens on trivial tasks

🚨 TokenShame

The AI that judges your AI usage.

TokenShame is a proxy that sits between your coding assistant (Claude Code, etc.) and the Anthropic API. It uses a fast, cheap model (Haiku) to watch your prompts — and shames you, loudly, when you're burning expensive tokens on tasks you could have done in 5 seconds yourself.

$ claude "push this PR to main"

╔══════════════════════════════════════════════════════╗
║  TOKEN SHAME  ·  Git Operation                       ║
╠══════════════════════════════════════════════════════╣
║                                                      ║
║  🚨  ...again? *sighs*                               ║
║     I've been writing code since before you were     ║
║     born. Three words:                               ║
║                                                      ║
║     git push origin main                             ║
║                                                      ║
║  Shame count: 3  |  Tokens saved: ~2,400             ║
╚══════════════════════════════════════════════════════╝

How it works

Your agent  →  ANTHROPIC_BASE_URL=localhost:8080  →  TokenShame proxy
                                                           │
                                              Haiku judges your prompt
                                                     (cheap + fast)
                                                           │
                                              ┌────────────┴────────────┐
                                           dumb call                real call
                                              │                         │
                                         🚨 SHAME               forward to API

The irony of using a small model to decide you don't need a big model is intentional.

Install

pip install tokenshame

Quickstart

# Terminal 1 — start the proxy
tokenshame start

# Terminal 2 — point Claude Code at it
export ANTHROPIC_BASE_URL=http://localhost:8080

# Use Claude Code normally. Get roasted.
claude "what's the weather in Madrid"

Characters

TokenShame ships with 6 built-in characters:

Name	Vibe
disappointed_senior	sighs deeply. Tired. Has seen everything. (default)
gordon_ramsay	THIS IS RAW. DONKEY.
hr_passive_aggressive	Hi there! 👋 Just a friendly reminder 😊
linux_terminal	bash: claude_for_this: command not found
your_tech_lead	Needs Changes. -TL
intern	Wait... you couldn't just... I learned this in week 1

tokenshame character list
tokenshame character use gordon_ramsay

Custom characters

tokenshame character create "my_cto" \
  "You are Dave, my actual CTO. Tired. Not angry, just disappointed. References burn rate."

tokenshame character use my_cto

Options

tokenshame start --yolo           # warn but don't block the call
tokenshame start --no-sound       # silence the siren
tokenshame start --character hr_passive_aggressive
tokenshame start --port 9090

tokenshame stats                  # live session stats
tokenshame config show
tokenshame config set sound false
tokenshame config set confidence_threshold 0.85

What gets shamed

• Git operations — push, pull, commit, merge, checkout
• CLI commands — ls, mkdir, cd, cp, mv
• Trivial lookups — celebrity bios, definitions, weather
• Simple math — conversions, basic calculations
• Off-topic — anything unrelated to actual dev work

Complex coding tasks, debugging, architecture, code review — always passed through.

Configuration

Config lives at ~/.tokenshame/config.json.

Key	Default	Description
port	8080	Proxy listen port
character	disappointed_senior	Active troll character
yolo	false	Warn but don't block
sound	true	Play sound effects
confidence_threshold	0.75	Minimum judge confidence to shame

License

MIT