Tax calculation you can prove.
A deterministic tax rules engine, delivered as an API. Every result ships with a signed, replayable calculation trace - each rule that fired, in order, citing the legislation it came from. Australia first; every G7 jurisdiction on one engine.
Same input, same output, forever. No model in the calculation path.
Legislation, compiled.
Tax law is already deterministic - the industry just hides it inside spreadsheets and seat-licensed suites. TaxKernel encodes each jurisdiction's fiscal year as a versioned, regression-tested ruleset and exposes it as infrastructure - same engine, same certificate format, from the ATO to the IRS.
One correct answer
Pure functions over encoded law. No sampling, no temperature, no drift. A calculation run today replays identically in a decade - which is what an audit requires.
Every number carries its trace
Responses include the full rule-firing sequence with provision-level citations, signed by the engine. Your users, auditors, and regulators can verify - not trust.
Law changes. History doesn't.
Rulesets are immutable per fiscal year and semantically versioned within it. Amendments ship as diffs against the statutes, tested against each revenue authority's published tables before release.
One call. A number and its proof.
Calculate anything the Acts define. Income tax, Medicare levy and surcharge, HELP repayments, offsets, CGT with discount, Division 293, payroll withholding.
Pin the year, pin the version. Request against au-2025-26 and the answer never changes underneath you.
Verify independently. The trace replays through the open verifier client-side - your compliance team doesn't have to take our word for it.
# POST /v1/calculate { "ruleset": "au-2025-26", "kind": "individual_income_tax", "input": { "residency": "resident", "taxable_income": 98000.00, "has_private_health": true, "help_balance": 0 } } # 200 - result + certificate { "net_tax": 22148.00, "trace": [ 3 rule firings, cited ], "signature": "ed25519:tk-prod-01:…" }
Jurisdictions
Australia is the flagship - full individual income tax and payroll withholding at launch. The same engine, rule format, and certificate then carry every G7 jurisdiction, each versioned to its own fiscal year and regression-tested against its revenue authority's published artefacts.
| Jurisdiction | Tested against | Status |
|---|---|---|
| Australia au-2025-26 | ATO withheld tables, simple tax calculator | live at launch |
| United Kingdom uk-2026-27 | HMRC rates & PAYE artefacts | next |
| United States us-2026 | IRS Rev. Proc. tables, Pub 15-T | next |
| Canada ca-2026 | CRA federal + provincial tables | roadmap |
| Germany de-2026 | BMF Programmablaufplan (official algorithm) | roadmap |
| France fr-2026 | DGFiP barème & official simulator | roadmap |
| Italy it-2026 | Agenzia delle Entrate IRPEF artefacts | roadmap |
| Japan jp-2026 | NTA withholding tax tables | roadmap |
The tax endpoint for AI agents
A language model should never do tax arithmetic. It should call something that cannot get it wrong.
Every AI assistant touching money - lending, payroll, advice, filing - needs a deterministic tax authority it can delegate to. TaxKernel ships as a native MCP server, so agents call it like a tool: structured input in, certified number out, trace attached for the human in the loop.
The certificate is the safety property. When an agent shows its user a tax figure, the figure carries proof of where it came from - not a probability of being right.
Pricing
Developer
free - 1,000 calcs/mo
- All live modules, current FY
- Full traces, unsigned
- Community support
Platform
from $499/mo + per-calc
- Signed certificates, open verifier
- All FY rulesets, version pinning
- MCP server access
- 99.9% SLA
Authority
annual licence
- "Calculated by TaxKernel" trust mark
- On-prem / VPC deployment
- Ruleset early access & diff briefings
- Compliance attestation pack