finance-skill（财务技能）

Finance Skill

Personal finance memory layer. Parse statements, store transactions, query spending.

Data Location

• Transactions: ~/.openclaw/workspace/finance/transactions.json
• Raw statements: ~/.openclaw/workspace/finance/statements/

*Storage convention: OpenClaw workspace (~/.openclaw/workspace/) is the standard location for persistent user data. This matches where session-memory and other hooks store agent data. Credentials/config would go in ~/.config/finance/ if needed.*

Tools

1. Parse Statement

When user shares a statement (image or PDF):

⚠️ IMPORTANT: Telegram/channel previews truncate PDFs! Always extract with pypdf first to get ALL pages:

python3 -c "
import pypdf
reader = pypdf.PdfReader('/path/to/statement.pdf')
for i, page in enumerate(reader.pages):
    print(f'=== PAGE {i+1} ===')
    print(page.extract_text())
"

Then parse the full text output:

1. Extract transactions from ALL pages
2. Return JSON array: [{date, merchant, amount, category}, ...]
3. Run scripts/add-transactions.sh to append to store
4. Verify total matches statement (sum of expenses should equal "Total purchases")

Extraction format:

Each transaction: {"date": "YYYY-MM-DD", "merchant": "name", "amount": -XX.XX, "category": "food|transport|shopping|bills|entertainment|health|travel|other"}
Negative = expense, positive = income/refund.

Categories:

• food: restaurants, groceries, coffee, fast food
• transport: Waymo, Uber, gas, public transit
• shopping: retail, online purchases
• bills: utilities, subscriptions
• entertainment: movies, concerts, theme parks
• health: pharmacy, doctors
• travel: hotels, flights

2. Query Transactions

User asks about spending → read transactions.json → filter/aggregate → answer

Example queries:

• "How much did I spend last month?" → sum all negative amounts in date range
• "What did I spend on food?" → filter by category
• "Show my biggest expenses" → sort by amount

3. Add Manual Transaction

User says "I spent $X at Y" → append to transactions.json

File Format

{
  "transactions": [
    {
      "id": "uuid",
      "date": "2026-02-01",
      "merchant": "Whole Foods",
      "amount": -87.32,
      "category": "food",
      "source": "statement-2026-01.pdf",
      "added": "2026-02-09T19:48:00Z"
    }
  ],
  "accounts": [
    {
      "id": "uuid",
      "name": "Coinbase Card",
      "type": "credit",
      "lastUpdated": "2026-02-09T19:48:00Z"
    }
  ]
}

Usage Flow

1. User: *shares statement image*
2. Agent: extracts transactions via vision, confirms count
3. Agent: runs add script to store
4. User: "how much did I spend on food?"
5. Agent: reads store, filters, answers

Dependencies

• jq — for JSON transaction storage and querying (apt install jq / brew install jq)
• pypdf — for full PDF text extraction (pip3 install pypdf)

Lessons Learned

• Telegram truncates PDF previews — always use pypdf to get all pages
• Verify totals — sum extracted expenses and compare to statement total before importing
• Coinbase Card — no Plaid support, statement upload only

Future: Plaid Integration

• Add finance_connect tool for Plaid OAuth flow
• Auto-sync transactions from connected banks
• Same query interface, different data source