Heath Ledger

AI bookkeeping skill for Mercury bank accounts.

Quick Start

1. scripts/init_db.mjs — creates DB + seeds ~90 universal vendor→category rules
2. scripts/connect_mercury.sh [entity_name] — discovers accounts
3. (Optional) scripts/connect_stripe.sh — connect Stripe for exact revenue + fees
4. (If Stripe connected) scripts/pull_stripe_revenue.sh — pull monthly revenue data
5. scripts/pull_transactions.sh
6. scripts/categorize.sh — rule-based first, AI for unknowns
7. Review ambiguous items, correct with scripts/set_category.sh
8. scripts/generate_books.sh [output_path]

Setup Flow

Mercury API Key (Required)

Get from Mercury Dashboard → Settings → API Tokens. The token gives read-only access to transactions.

Stripe API Key (Optional but Recommended)

Without Stripe API: Mercury shows net Stripe deposits (revenue minus fees). The system estimates gross revenue using a configurable fee rate (default 2.3% + $0.30).

With Stripe API: You get exact gross revenue, exact fees, and proper refund tracking. Always prefer this when available.

To connect: scripts/connect_stripe.sh

Then pull data: scripts/pull_stripe_revenue.sh

The P&L generator automatically uses Stripe data when available, falling back to Mercury estimates otherwise.

Entity Settings

Configure per-entity via the entity_settings table:

Workflow

1. Connect Mercury — scripts/connect_mercury.sh [name] discovers accounts, creates entity
2. Pull transactions — scripts/pull_transactions.sh
3. Categorize — scripts/categorize.sh [max_transactions] — rule-based first, then AI for unknowns
4. Review ambiguous — Script outputs low-confidence items. Ask user, then update with scripts/set_category.sh [subcategory]
5. Generate books — scripts/generate_books.sh [output_path]

Scripts Reference

All scripts are in scripts/. Run with bash or node. Database is SQLite at data/heath.db.

Chart of Accounts

See references/chart-of-accounts.md for the full chart with P&L sections and cash flow classifications.

Learning & Compounding System

Heath Ledger gets smarter over time through a layered rule system:

Rule Hierarchy

1. Entity-specific rules (highest priority) — per-company overrides
2. Global rules (entity_id = NULL) — apply to all entities
3. Seed rules — universal vendor mappings shipped with the skill
4. AI categorization — used when no rule matches

How Learning Works

• Every manual correction creates or updates a categorization rule
• Rules track usage_count — heavily-used rules are more reliable
• source field tracks provenance: seed, ai, human, manual
• Human-confirmed rules get confidence: 0.95-1.0
• AI-generated rules start at 0.85 and can be promoted
• Entity-specific rules can be promoted to global when they prove universal

The Compounding Effect

After categorizing ~5,000 transactions across 2 entities, the system now auto-categorizes ~95% of transactions without AI. Each new entity benefits from all previous learnings.

Known Limitations

Stripe Net vs Gross (Without Stripe API)

Mercury deposits from Stripe are net amounts (revenue minus ~2.9% + $0.30 fees). Without the Stripe API:

• We estimate gross revenue using configurable fee rates
• This creates "synthetic" Stripe Fee entries
• Accuracy depends on your actual Stripe fee rate (varies by plan, card type, international)
• Solution: Connect Stripe API for exact numbers

Deel Fee Splitting

Deel combines platform fees and contractor payroll in one transaction stream. Pattern:

• Small fixed amounts (~$2-5) → Deel Platform Fee → categorize as "Software expenses"
• Larger variable amounts → Contractor Payroll → categorize as "Wages & Salaries"
• The system learns this pattern but may need initial human guidance

Mercury API Limitations

• Only returns posted transactions (not pending)
• Some counterparty names are truncated or normalized differently
• Wire descriptions may include reference numbers that create duplicate rules

Multi-Currency

• Wise transfers create both a debit (USD) and may show FX fees separately
• International wire fees from Mercury appear as separate line items
• FX gains/losses are not tracked (would need multi-currency ledger)

AI Categorization

The categorize.sh script calls the host agent's model via stdin/stdout JSON protocol. It sends transaction batches and expects category assignments back. The script writes a prompt to stdout that the agent should process and return results for.

When AI confidence < 0.85, transactions are flagged as ambiguous for user review.

Key Details

• Cash or accrual basis — configurable per entity
• Multiple entities supported — each with own connections and rules
• Rules persist — categorization rules saved to SQLite, reused across runs
• Seed rules — ~90 universal vendor mappings loaded on init
• Excel output — 4-tab workbook: P&L, Balance Sheet, Cash Flow, Transaction Detail