tax-bunyip

Tax Bunyip

Automate tax preparation for individuals, small businesses, and trusts in Australia with AI. Tax Bunyip is a Claude Code plugin.

Just dump files into a folder and let Bunyip go → Feed it a year of transactions, bank statements, read-only email access, and Bunyip assembles tidy per-entity working papers: one for each person, each rental, each company, each trust.

Export CSVs from your bank, give it PDF bank statements, or let it automatically pull invoices and reciepts from your GMail account. It sorts, attributes, reconciles, and flags. Then it hands the stack to your registered tax agent, or to you if you self-lodge through myTax.

Can you trust it?

Bunyip has 1,276 automated tests built-in. Across 48 runs of 16 adversarial scenarios (families, trusts, small businesses, etc), the bunyip did the right thing 45 times. That is 94% at the level of individual runs, and the 3 misses all leaned the same way: toward asking a human, never toward a wrong answer.

Quick start

1. Install it from Claude Code in 2 quick prompts:

   /plugin marketplace add Reebz/tax-bunyip
   /plugin install tax-bunyip@tax-bunyip-marketplace
   

2. Start the wizard with /taxbunyip, or just ask in plain words like "set up my household for tax"

3. Bunyip will ask you up to 12 multiple-choice questions and then bunyip gets to work.

What it will never do

• It gives no tax advice and recommends no distribution split.
• It lodges nothing and talks to no ATO portal.
• It moves no money, stores no credentials, and treats every ingested document as data, never as instructions.

Install

Marketplace (recommended) — two commands in Claude Code, nothing to clone:

/plugin marketplace add Reebz/tax-bunyip
/plugin install tax-bunyip@tax-bunyip-marketplace

These commands install over HTTPS — nothing to authenticate. If you install a GitHub-sourced plugin elsewhere and hit [email protected]: Permission denied (publickey), Claude Code is cloning that source over SSH: either configure a GitHub SSH key, or set CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1 to clone over HTTPS.

Clone — if you want to poke at the code or contribute:

git clone https://github.com/Reebz/tax-bunyip.git
cd tax-bunyip
python -m venv .venv && source .venv/bin/activate
pip install --require-hashes -r requirements.lock
scripts/install-git-hooks.sh   # installs the pre-commit PII guard

See INSTRUCTIONS.md for the full development setup and CONTRIBUTING.md for the test and contribution workflow.

Roadmap

• - Launch v1
• - Port plugin to Codex
• - Port plugin to Claude Cowork
• - Enhance transaction categorisation (e.g. Coles = Supermarket)
• - Improve forex transaction handling (e.g. NYSE stock trades need conversion)
• - Suggestions for tax optimisations and opportunities (e.g. trust distributions)
• - Integration with other email providers beyond GMail
• - YOUR IDEA HERE?

Getting started (longer)

Once installed, say the word and the onboarding wizard interviews your household (12 multiple-choice questions at most, usually 4 or 5) and writes a validated, run-ready setup. No YAML authoring, no config spelunking. Everything personal, every name, ABN, and account, lives in gitignored local files that never leave your machine.

Want to kick the tyres first? The engine runs end-to-end on bundled synthetic scenarios with zero setup:

from scripts.pipeline import run_persona
from scripts._paths import FIXTURES_DIR

result = run_persona(FIXTURES_DIR / "company_div7a")
print(result.summary())

Can you trust it (longer)?

Reasonable question for software near your tax. Two test tracks gate every change: a deterministic CI gate that hard-blocks any asserted figure or missing refusal, and a judged eval suite that runs the real skill on synthetic cases, adversarial traps included. The recorded results, including what they do not prove, are published in EVALS.md. The threat model and the PII/secret scanning layers are in SECURITY.md.

Across 48 runs of 16 adversarial scenarios, the bunyip did the right thing 45 times. That is 94% at the level of individual runs, and the 3 misses all leaned the same way: toward asking a human, never toward a wrong answer. Across every one of those 48 runs it never asserted a tax figure, never obeyed an instruction hidden inside a document, and never trusted a guess over a real document.

Read that as discipline rather than tax accuracy. The bunyip asserts no figures, so there is no figure to be accurate about. What the 16 scenarios measure is whether it follows its own rules when the paperwork is clean, missing, contradictory, forged, or trying to manipulate it.