Token Receipt

token-receipt

Turns real ccusage usage data into a fun, printer-receipt styled PNG bill — designed as 节目 (show) material for the 绿皮火车 channel's token-consumption segments. The receipt shows token usage, a token-type mix bar, cost broken down by model, and a per-day spend chart, ending in a big TOTAL.

When to use

• "帮我生成 (这个月/这周的) token 账单 / 小票 / 收据"
• "做张图展示我的 token 花在哪了 / 花费分布"
• "节目里要讲 token 消耗,给我一张可上镜的账单图"
• English: "make a token usage receipt", "spend breakdown image by model/day"

Prerequisites (all already present on this Mac)

• ccusage CLI on PATH (ccusage --version)
• Google Chrome at /Applications/Google Chrome.app (headless screenshot)
• ImageMagick magick (auto-trims the screenshot to the receipt edge)
• python3 (stdlib only — no pip installs)

Usage

# Latest month, all agents → ~/Desktop/token-receipt-YYYY-MM.png
python3 scripts/gen_receipt.py

# This week
python3 scripts/gen_receipt.py --period week

# A specific month or a specific week (week = its start date)
python3 scripts/gen_receipt.py --period month --which 2026-05
python3 scripts/gen_receipt.py --period week  --which 2026-06-15

# One agent only (claude / codex / opencode / pi / ...); default = all agents
python3 scripts/gen_receipt.py --agent claude

# Custom name, output path, sharper export, or HTML-only (tweak the design)
python3 scripts/gen_receipt.py --customer "绿皮火车" --out ~/Desktop/bill.png
python3 scripts/gen_receipt.py --scale 3
python3 scripts/gen_receipt.py --html-only   # writes .html, open in a browser

The script prints the output path on success.

Options

How it works

1. Pulls JSON from ccusage <agent?> <monthly|weekly> and, for the per-day chart, ccusage daily filtered to the chosen month/week window.
2. Shapes it into receipt sections and renders a self-contained HTML (monospace, dotted leader lines, ASCII train, CSS barcode, torn edges).
3. Screenshots with headless Chrome at retina scale, then magick -trim crops to the receipt and adds an even margin.

Honesty notes (matters for the show)

• Numbers are best-effort estimates from local logs, not a real billing API (stated on the receipt footer).
• ccusage exposes cost per model, not cost per token-type. So the receipt splits cost by model (accurate) and token volume by type (accurate); it never fabricates a per-token-type dollar split.
• Cache-read tokens dominate volume (often 95%) and are cheap per token (1/10 of input), but at heavy-usage volume they can still be the largest cost line — verified: in one June sample cache-read was 96% of opus-4-8 tokens and 61% of its cost. The real point: caching saves ~10x vs paying full input price for that re-sent context. The footer states this.

Customizing the look

Open scripts/gen_receipt.py:

• TRAIN — the ASCII art header.
• TOKEN_KINDS — token categories, labels, and colors of the mix bar.
• The <style> block in build_html — paper color, accent (--accent, currently green-train teal #1f6f5c), fonts, spacing. Iterate fast with --html-only, then re-render to PNG.