Finance Assistant

Real math. Real law. Your machine.

A personal finance copilot that applies actual tax statute to your real numbers — not estimates, not vibes. Runs locally through Claude Code; your data never leaves your machine. Open source, free forever.

6 locales · 14 bank formats + any file via LLM · Monte Carlo FIRE · 1,400 tests

→ Live demo — see a real conversation without installing

Applies real tax statute to your numbers, but it's software, not a licensed advisor — for filing decisions, confirm with a tax professional or your local authority.

See it work — real output, not a mockup

FIRE confidence, with the actual distribution:

You:  What's my FIRE confidence level?
FA:   Running 10,000 simulations (return 7%±12%, inflation 2%±0.8%)...

      Retirement probability: 73% success at 2042 target
      ┌─────────────────────────────────┐
      │  p10   $420k  ████             │
      │  p25   $580k  ██████           │
      │  p50   $790k  ████████         │  ← median
      │  p75  $1.05M  ██████████       │
      │  p90  $1.38M  █████████████    │
      └─────────────────────────────────┘
      Current: $180k  target: $900k

      +$200/month raises success to 81%.

"Should I go freelance?" — with real SE tax + QBI, not a guess:

You:  I make $80k as a W-2 employee. What changes if I go 1099 at the same $80k?
FA:   Same $80k, two tax pictures (2024, single, federal):

      As a W-2 employee:
        Federal income tax        $9,441
        Your FICA (SS + Medicare) $6,120   (employer pays another $6,120)

      As a 1099 freelancer:
        Self-employment tax       $11,304  (you pay BOTH FICA halves — Schedule SE)
        − ½ SE tax deduction      −$5,652
        − §199A QBI deduction     −$11,950 (IRS Rev. Proc. 2024-40)
        Federal income tax        $5,569
        Total federal + SE        $16,872

      → Freelancing at the SAME number costs ~$1,300 more, because SE tax
        replaces the employer's FICA share. Break-even ≈ $82k. Want the rate?

Every figure is deterministic Python applying real statute — and it's validated against 33 official tax-authority test cases. More examples ↓

Who it's for

People who want correct answers about their own money — not a chatbot's plausible guess and not a SaaS that wants your bank login. Especially:

• Anyone with real tax complexity — freelance/1099, multiple countries, equity, "what if I change X" questions — tired of estimates.
• Privacy-conscious people who want their financial data to stay on their own disk.
• Claude Code users who'd rather talk to their finances than fill in a dashboard.

If your country isn't covered, the scaffold generator makes adding it a ~1-evening contribution.

Why this over a general AI assistant?

Any assistant can talk about money. The difference is law-accurate, bracket-correct tax math for six countries, computed by deterministic code — not a model guessing brackets. A general (or first-party) assistant gives you plausible estimates; this applies the actual statute (German EStG §32a, IRS Rev. Proc., HMRC PTM) to your real numbers, shows its work, and is tested against official authority cases.

Install

One line — clone straight into your skills folder (Claude Code auto-discovers it, no config to edit):

git clone --recurse-submodules https://github.com/googlarz/finance-assistant.git \
  ~/.claude/skills/finance-assistant
pip install -r ~/.claude/skills/finance-assistant/requirements.txt

Start a new Claude Code session and ask: "What's my financial health?"

python3 ~/.claude/skills/finance-assistant/skill.py --demo     # try it with sample data first
python3 ~/.claude/skills/finance-assistant/skill.py --doctor   # verify your setup

cd /path/to/your/finance-assistant
python3 skill.py --install     # symlinks into ~/.claude/skills/ (git pull stays live)

What it does

A conversational copilot across the full personal-finance lifecycle. The highlights:

• Tax, law-accurate — deductions, filing brief for your accountant, and saveable what-ifs (W-2 vs 1099, single vs MFJ, max the 401k?) computed from real brackets in 6 countries.
• Plan & project — budgets, savings goals, debt avalanche/snowball, FIRE with 10,000-path Monte Carlo, rent-vs-buy and salary-offer comparisons.
• Track wealth — portfolio allocation/XIRR, live stock + crypto prices, net-worth health score.
• Import anything — 14 bank-CSV formats on a fast path; an unrecognized file (foreign bank, scanned PDF, screenshot) is read by Claude directly, same sanitize→dedupe→confirm pipeline.
• Stay ahead — every session surfaces only what needs attention: budget overspend, upcoming bills, tax deadlines, portfolio drift, zombie subscriptions ("still charging after you flagged it to cancel"), FIRE progress. Stale alerts auto-suppress. Say hi for a one-line proactive check-in — it picks the single most relevant thing and asks you about it.
• Own your data — local SQLite, encrypted backup, append-only audit log of every change, one-command total wipe.

More conversations

You:  hi
FA:   Hey — you've paid off 50% of the credit card. That's the halfway mark.
      Sticking with the current plan or want to revisit?

You:  Best strategy to pay off my 3 loans?
FA:   Avalanche (highest rate first):
        Credit card $3,200 @ 18.9% → Car loan $8,600 @ 5.9% → Personal loan $13,000 @ 3.4%
        Interest saved vs minimum: $2,840 | Debt-free: Nov 2027
      Snowball (smallest first): saves $1,920 | Debt-free: Feb 2028
      → Avalanche saves $920 more and finishes 3 months earlier.

You:  How am I doing on my April budget?
FA:   $340 of $400 food (85%) — 28 days into 30-day month. On pace.
      Eating out: $89 / $80 — OVER by $9 ⚠
      Groceries has $60 remaining — cooking covers the gap.

The tax depth isn't US-only — ask "what can I deduct?" in Germany and you get German law (Arbeitnehmer-Pauschbetrag, Homeoffice-Pauschale, EStG §9a), not a translation of US rules. UK pulls HMRC pension carry-forward; France applies the quotient familial. Six countries, one set of commands.

Locales & validation

Tax rules live in country-specific plugins (a git submodule) that the skill loads dynamically.

All six are validated against 33 official tax-authority test cases (BMF, HMRC, DGFiP, Belastingdienst, KAS, IRS) — python3 -m pytest locales/tests/test_validation.py -v. US cases are computed independently from IRS Rev. Proc. 2023-34 and Schedule SE. Adding a country is independent of the main code — see the locales repo.

Where this runs

Built for Claude Code (the free CLI). A claude.ai Projects template ships too, but the browser sandbox can't reach your files, so most features need the local install.

Want zero data to Anthropic? For full privacy, run Claude Code against a local model: point claude-code-router at Ollama and nothing leaves your machine — not even the conversation. The deterministic tax math is identical; only the reasoning quality drops. Recipe + an accuracy harness you run yourself: docs/sovereignty.md.

Privacy & security

Local-first by design. Your data lives only in .finance/ on your machine — no cloud, no telemetry, nothing uploaded.

• Encrypted at rest — Fernet (AES-128-CBC + HMAC-SHA256), PBKDF2 480k iterations, atomic writes.
• Hardened — .finance/ is chmod 600/700 and auto-added to .gitignore on first run.
• You own the delete button — wipe any category or everything, one command.
• Audit log — every change is recorded; --audit shows what happened.
• Never stored — bank credentials, IBANs, card numbers, government IDs, raw documents.

This protects your data at rest. By default Claude Code still sends the conversation to Anthropic's API — for zero egress, route it through a local model (Sovereignty mode).

Full detail — encryption parameters, threat model, all controls, known limitations — in docs/SECURITY.md.

Docs

• Architecture — how it works, data layout, full module reference, testing
• Security — encryption, threat model, controls
• Sovereignty mode — run fully local via Ollama + accuracy harness
• Contributing — add a locale (~7 files), plugin spec, provenance format

Troubleshooting: ModuleNotFoundError: No module named 'locales' → you cloned without submodules; run git submodule update --init --recursive, then python3 skill.py --doctor.

Open source under CC BY 4.0. Built on real statute, tested against official numbers, run on your own machine.