Snap a photo of any receipt or invoice. Get structured data in your Google Sheet.

Photo → Gemini Vision → Structured JSON → Google Sheets row

Zero infrastructure. Works as a CLI tool, a hermes-agent skill, or a Telegram bot integration.

git clone https://github.com/lmanchu/receipt-scan.git
cd receipt-scan
npm install
cp .env.example .env
# Set GOOGLE_SHEETS_ID in .env

# Scan a receipt
node scan.js receipt.jpg

# Dry run (extract only, don't write to sheet)
node scan.js receipt.jpg --dry-run

# JSON output
node scan.js receipt.jpg --json

# From stdin
cat receipt.jpg | node scan.js --stdin

Output:

Merchant: 7-ELEVEN
Date:     2026-04-09
Amount:   TWD 89
Category: groceries
Items:    茶裏王, 御飯糰
Tax:      4

✅ Appended to Google Sheet

1. Image sent to Gemini 2.5 Flash Vision (via CLIProxy or any OpenAI-compatible endpoint)
2. LLM extracts: merchant, date, amount, currency, items, tax, category, invoice number
3. Structured data appended as a new row in your Google Sheet
4. Headers auto-created on first run

Install as a hermes skill:

ln -s $(pwd) ~/.hermes/skills/receipt-scan

Then send any photo to your Gloria/Hermes Telegram bot. The agent will detect receipts and offer to scan them.

• Store receipts (any language, any currency)
• Taiwan 統一發票 (electronic and paper)
• Restaurant bills
• Invoices and PDFs
• Uber/Grab/taxi receipts
• Credit card statements
• Handwritten receipts

See .env.example for all options. Key settings:

• GOOGLE_SHEETS_ID — your spreadsheet ID (required)
• LLM_MODEL — vision model (default: gemini-2.5-flash)
• BASE_CURRENCY — your currency (default: TWD)

• Node.js 18+
• Google OAuth token (reuses hermes-agent/Gloria token)
• Vision LLM access (Gemini via CLIProxy, OpenAI, or any compatible endpoint)

TaxHacker — self-hosted AI accounting app. We took the concept (LLM receipt extraction) and made it zero-infrastructure: no Docker, no PostgreSQL, just a script that writes to Google Sheets.

MIT