Skip to content

Lodgeit Labs · Agentic accounting

AI that calculates. It doesn’t guess.

Your AI assistant is good at reading, sorting and explaining. It should not be doing the arithmetic on a tax return. Lodgeit gives it a fleet of calculators and tax rules written as code, so every figure is computed, sourced and repeatable, and a person approves anything that matters.

See what’s liveConnect the calculators
  • No language model in the calculation path
  • 23 calculators live, free to call
  • Rules reviewed by named practitioners

The rule we build to

Each part does the job it’s good at.

A language model can sound right and be wrong by a few hundred dollars. So we split the work three ways and let systems enforce the split.

01

AI reads, routes and explains

It reads the client’s documents and ledger, works out which calculation applies, gathers the facts and explains the result in plain English.

02

Engines compute and prove

Calculators and rule files produce the figures. The same inputs always give the same answer, and each answer names the rate tables or rules it used.

03

People approve

Agents draft and stage. An accountant confirms anything that reaches a client, changes a ledger or goes to the ATO.

Guess or calculate

The same question, answered two ways.

A language model on its own
A model calling Lodgeit engines
Ask twice
Figures can differ between runs and between models.
The same request, period and rate tables return the same response, byte for byte.
Where the number came from
A paragraph of reasoning you have to take on trust.
Every response lists the statutory rate tables it used, each with a SHA-256 content hash.
Outside its scope
Answers anyway, with a plausible number.
Refuses, and says why in a form an agent can act on.
Rounding
Decided afresh each time.
One published policy, named in every response. Money comes back as exact decimals.
Who checked the logic
Nobody you can name.
Practitioners whose verdicts sit on a public registry, tied to the version they reviewed.

What’s live today

Four services your assistant can reach.

Each connects through the Model Context Protocol, so it works with Claude and other assistants that support connectors. Bring your own agent.

CalculateLive · open access

The calculator constellation

Deterministic Australian tax and accounting calculators behind one public gateway. No key, no sign-in, no cost. REST and MCP return the same body.

19 Fringe benefits tax1 Division 7A2 Depreciation (AASB 116)1 Hire purchase
  • FBT: car, loan, debt waiver, expense payment, housing, LAFHA, board, entertainment, car parking, property and residual benefits, with the grossed-up figures and workings.
  • Division 7A: minimum yearly repayment, shortfall and deemed dividend, on the benchmark rate for the year.
  • Depreciation and hire purchase: carrying amounts at any date or range, and full amortisation schedules with the current and non-current split.

Calculator output is not advice. Response shapes are not yet versioned for production integration.

Calculator overviewllms.txt for agents
ResearchLive

Tax Genii

An Australian tax assistant that retrieves legislation and ATO material and answers with citations. Over MCP it offers seven tools, from raw legislation search to a reviewed multi-agent analysis.

  • Rules where they exist. For Division 36 tax losses of resident individuals, the figures come from Lodgeit’s rules engine, and the response identifies the rules file.
  • Asks before it assumes. If the rules need a fact such as residency, it asks for it.
  • Declines out of scope. A company loss question is declined by the rules and answered as AI reasoning, labelled as such.

Under reviewThe Division 36 rules are under practitioner review. Everything else is AI reasoning over retrieved law, with citations.

Try Tax Genii
ProveLive

SBRM, the accounting engine

Give it a general ledger, a chart of accounts and an opening trial balance. It builds the four primary statements and proves them with six checks, each pass or fail with a dollar variance.

  • Loads your exports from Xero, QuickBooks Online and MYOB, and compares its trial balance with the source, account by account.
  • Explains a failed check and verifies depreciation and hire purchase schedules against the ledger.
  • Corrections are previewed first, applied only to SBRM’s copy after you agree, and can be undone.
The six checksAbout SBRM
EngageLive

ClientRelay

The practice front door: proposals, e-signing and payments, client forms, and KYC/AML checks. Connect your assistant and it works inside your firm’s own templates and service list.

  • Drafts proposals from your catalogue, such as next year’s renewals for a whole client group.
  • Builds forms from a description, including uploads and show/hide rules, from reusable blocks.
  • Sees due diligence status, so it can tell you who is verified and what is still missing.

It drafts. Nothing reaches a client until you press send.

Engage & get paidClient formsKYC / AML

Refusals are a feature

A refusal you can see beats a number you can’t check.

Ask the accounting depreciation calculator about a pooled asset and it won’t compute one under the wrong rules. It returns a typed refusal instead.

  • Status codes mean what they say. 422 is a malformed request and names the field. 400 is a deliberate refusal. A 5xx is our fault.
  • Elections stay with the taxpayer. Where the law offers a choice of method, agents are told to compute each one the records support and let the employer choose.
  • Every figure says what it is. An AASB 116 carrying amount carries a warning against using it in a tax return.
depreciation-at · HTTP 400
// request: a $10,000 asset flagged as a low-value pool asset
{
  "detail": {
    "refusal_class": "pool_asset_out_of_t6_scope",
    "refusal_payload": {
      "pool_type": "low_value_pool",
      "asset_acquisition_date": "2023-07-01"
    },
    "engine": "depreciation",
    "message": "depreciation refused the request:
                pool_asset_out_of_t6_scope"
  }
}

Prove the books

Six checks on every set of statements.

SBRM is Lodgeit’s deterministic accounting engine, built on Charlie Hoffman’s Standard Business Report Model. An agent can run it before and after any adjustment, so a change that breaks the books shows up straight away.

  • No plugs. Our own pipeline fails a run if a movement is filled in without a cited source line.
  • A result that looks too clean is a warning. Unexplained amounts are reported and classified, never smoothed over.
CheckWhat it provesResult
Accounting equationAssets equal liabilities plus equityPass · $0.00
Equity roll-forwardOpening equity plus profit and other movements equals closing equityPass · $0.00
Cash flow roll-upThe cash flow sections add upPass · $0.00
Cash roll-forwardOpening cash plus net cash flow equals closing cashPass · $0.00
Cash consistencyClosing cash in the cash flow equals cash on the balance sheetPass · $0.00
Groups add upEvery group’s total equals the sum of what is in itPass

Illustration of the output shape. Each check returns pass or fail with its dollar variance.

Rules as code, grown with the profession

We write the law out as rules. Practitioners tell us where we’re wrong.

Each engine starts as statute written out as rules and tests. Before we rely on it, we pay people who have applied that law in practice to try to break it.

funded review bounties so far
15
total bounty pool
$13,500
typical effort per review
1.5–3 hrs
paid within, once a verdict is accepted
30 days
01

Choose a bounty

Each is one bounded statutory question. Email to reserve it before you start.

02

Read the brief

A worked example against the engine’s output, a set of edge cases, and the statute already cited.

03

Give a verdict

Accept, reject with reasons, or fix with the correction. Cite where you differ.

04

Your name on the record

Accepted verdicts go on a public registry with a hash tying your review to the version you checked. Initials or anonymous if you prefer.

Reviewed and paid so far

Div 7A minimum yearly repaymentDepreciation effective lifeSmall business general poolFBT car: statutory vs operating costFBT car parkingHire purchase interestCGT 15-year exemptionCGT active asset testPSI attribution

Open or in review now

CGT event A1 and net capital gain, in five partsCoal LSL levyDivision 36 loss sequencing, revision 2Instant asset write-offLegal opinion on rewarding rule signers

For CA, CTA, CPA, IPA members and registered tax agents. Lapsed membership and retirement are both fine.

See the bountiesHow we get quality

Where this is heading

A marketplace of calculators and rules, not one vendor building everything.

Australian tax law is too wide for any one company to encode. So the constellation is built to hold engines we didn’t write, as long as they meet the same standard.

Builders publish

Build a bounded, deterministic tax or accounting calculator and publish it into the constellation. Host it yourself and we point at it, or we run your engine behind the shared gateway.

Read the publishing standard

Practitioners attest

Reviewers check the statute-to-code translation and sign their verdict. Their name travels with the engine they reviewed.

From bounties to ownership

Firms and agents use

One registry lists every calculator, the periods it accepts and its input schema. An agent discovers what exists and calls it, without a custom integration for each.

How agents discover them

Live

The standard

A calculator must describe itself, refuse what it can’t answer, name what it consumed and return money as exact decimals.

Live

Conformance-passed

The running code is the reviewed code, and its numbers reproduce against a pinned source anyone can inspect.

In progress

Practitioner-attested

A practitioner has reviewed the logic against the statute. A passing bounty verdict lifts an engine to this tier.

Planned

Paid per use

Per-call payment and a signed public registry are designed but not live. We’re taking legal advice on rewarding practitioners when a rule they signed is used.

First external engine: a Coal Mining Industry long service leave levy calculator, built and hosted by contributor Ryan Duguid. It is conformance-passed and waiting on practitioner review.

Across the whole engagement

From first contact to a proven set of books.

The same pattern runs through the practice, with your assistant drafting and your team deciding.

Engage

Proposal, signature, payment

Your assistant drafts the proposal from your service list. The client signs, Stripe collects on schedule and each invoice lands in QuickBooks Online.

Collect

Only the questions that apply

Forms built from reusable blocks, with show/hide rules and uploads. Free and unlimited on every ClientRelay plan.

Verify

Know who you’re acting for

Identity, company and trust checks at about $5.50 a person. Possible matches are shown for your review. The decision stays yours.

Calculate and prove

Figures with their workings

Calculators and rules produce the numbers, and SBRM checks the statements still hold together.

KYC price excludes GST and is billed by our verification partner, Admiin. ClientRelay supports your due diligence. Your AML/CTF program and decisions remain your responsibility.

Being built

What’s next, and how far along it is.

  • Built, not yet live

    Tax depreciation calculators

    Decline in value under Division 40, with the low-value pool and the small business pool. Built and passing its tests, waiting on release.

  • In review

    Capital gains rules

    CGT event A1, cost base, the main disregards, the discount and the net capital gain method statement, written as rules and open for practitioner review.

  • In development

    Financial report review

    An AI reads a financial report, statements and notes, and SBRM proves it. Every figure is cited to its position in the document and nothing is filled in.

  • In development

    Year-end adjustments, staged for approval

    Agents read the books, calculate the adjustment, prove the result and stage the journal. An accountant sees the before and after and confirms before anything is posted.

  • Planned

    Agent skills for year-end jobs

    One skill per job, starting with the income tax provision, then accounting depreciation, hire purchase, Division 7A and FBT. Each tested against a client file with a known answer.

Get started

Three ways in.

Accounting firms

Add the calculators to Claude as a custom connector under Settings, Connectors. No sign-in is needed. Then ask it a Division 7A or FBT question and check the figure against your own.

Connector steps and current URL

Watch: using Claude with ClientRelay

Developers and agent builders

Give your agent one file. It lists every endpoint, the error contract, the number formats and the known limits, and points to the full OpenAPI schema.

lodgeit.org/llms.txt

Publish your own calculator

Tax practitioners

Pick up a review bounty. A couple of hours on a section you know well, paid on acceptance, with your name on the public record.

Open bounties

Reserve one by email

FAQ

Questions practices ask.

Can’t find the answer? Contact support.

Is the calculator output tax advice?

No. Every response says so and names the statute it relies on. The calculators compute from the inputs you give them and don’t judge whether those inputs are reasonable. A registered tax agent should review any figure before it is relied on.

Is Tax Genii deterministic?

Only in one area so far. For Division 36 tax losses of resident individuals, the figures come from Lodgeit’s rules engine, which is under practitioner review, and the response identifies the rules file. Everything else is AI reasoning over retrieved law, with citations you should read at the source.

Can an agent post to my ledger or send things to my clients?

No. Agents draft and stage. Proposals and forms stay as drafts until you send them, and journal posting, when it arrives, will need an accountant to confirm each entry.

What does it cost to use the calculators?

Nothing. There is no key, no sign-in and no charge. Each service runs a small number of instances, so sustained heavy load will queue. Talk to us first if you need more than that.

Can I build the calculators into my own product?

You can call them today, but response shapes are not yet versioned and the gateway address will change, so keep the address in configuration and expect changes. Get in touch before depending on them in production.

Which AI assistants does this work with?

The services use the Model Context Protocol, so they work with Claude and with other assistants that support MCP connectors. The calculators are also plain HTTP for anything else.

What if I think a number is wrong?

Send us the request, the response, the figure you expected and where that expectation comes from. Independent re-runs have found most of the defects fixed so far, and corrections are published with their date.

Give your AI something to calculate with.

Connect the calculators in a couple of minutes, or help us check the next rule.

Connect the calculatorsReview a rule
Back to top