U.S. tax return App

Snap your tax documents. Get your federal return, explained in English, Spanish or Chinese. Read every line of the code behind it.

Try Libertas →

  • 345,700 lines of code, written by AI
  • ~100,200 lines of tax calculation
  • ~36,800 lines of input validation
  • ~49,400 lines of explanations, scope and record handling
  • 190 different full screens, including each document editor
  • 7,911 fields a filer can see
  • 12,099 fields in the public schema
  • 59,317 localized strings
  • 11,324 runtime messages
  • 1,107 lines of display rules

Read the metrics

AI writes as much as it can.
People maintain as little as they must.

Where the code goes

Most of the App is tax law. Screens, storage and the connection to Libertas together take 0.7% of the code.

What the App needs Lines Share
Tax calculation, validation and explanations 186,455 53.9%
Schema and data declarations 61,131 17.7%
Interview, navigation and page mapping 58,631 17.0%
Localized runtime messages 30,698 8.9%
Return persistence 1,153 0.33%
Display rules for 190 screens 1,107 0.32%
Libertas platform integration 154 0.04%

Reference tables, wiring and shared helpers make up the remaining 1.9%.

Libertas packs the App's information into very few lines:

  • Platform integration: 154 lines. Eleven functions form the App's boundary with Libertas: they take each request from the Hub and read and write the return through the Data API. Screens, AI Import, translation and access control come from the platform.
  • Calculation and validation. Of the 186,455 tax lines, about 100,200 calculate the return and about 36,800 check the facts entered. The other 49,400 or so explain each figure with its form line and IRS source, handle what is out of scope, and select and convert records. Calculation helpers in the interview code add about 7,000 lines, for about 107,300 lines of calculation in all.
  • Schema fields: 20,346 declared. They take 61,131 lines, internal records included: about three lines per field with its type, limits and checks. The public schema has 12,099 fields, 7,911 of them visible to a filer. Labels and help are doc comments, outside the line count.
  • Localized strings and runtime messages: 59,317 strings. Every label, help text and message can be translated. 11,324 of them are runtime messages, form line labels included, that the code selects as it runs, many with placeholders such as a note's amounts. They take 30,698 lines.
  • Screens: 190, arranged by 1,107 lines. Every screen comes from the declarations; display rules arrange them, under six lines a screen. The interview code decides which questions apply and converts answers to stored records; it draws nothing.

Counts are from the metrics report, excluding tests. Calculation and validation lines are estimates from reviewed random samples.

Every figure explains itself

Every line the App calculates comes with its explanation: how the amount was worked out, the IRS form, instructions or publication it follows, and the entries and earlier lines it came from.

When the App can't calculate something, it doesn't guess. Each gap is its own note on the Review page. The note says what is out of reach and links to the page where you enter the missing fact. Nothing silently counts as zero.

Document center: your paperwork is the interview

Most tax software makes you retype your forms. Here you start with the documents themselves. Photograph or upload your W-2s, 1099s, K-1s and other statements, and AI reads each one into the matching entries. You check the result and save it; nothing is accepted without you. Each document stays listed under whose it is, and every figure on the return traces back to it.

The interview then asks only about what your documents cannot say: your household, your dependents and your situation.

Written by AI. Open line by line.

All of it is public on GitHub. Every rule, limit and calculation is source code you can read, with the IRS form or publication it comes from.

Coverage: TurboTax-level, and closing fast

The 80/20 rule: 80% of the code serves only 20% of what filers need. Wages, interest and the standard deduction take a small part of it; farms, rentals, K-1s and foreign income take the rest. AI writes that long tail, so one App can aim for the coverage of TurboTax. Changes land every day.

  • Federal, today. Form 1040 with its schedules, worksheets and credit forms: wages, business, farm, rental, investments, retirement, foreign income, education, health coverage, and much more. The coverage audit lists every form and its limits.
  • State, next. State and local tax is not calculated yet. The state boxes on your W-2s and 1099s are already kept for it.

It is a 2026 estimate on the IRS draft forms and tables, with synthetic data. It does not print or file a return.

Try it with sample papers

No retyping. Hand the App the paperwork and watch AI read it.

Two synthetic households are ready for you:

  1. Casey Sample: simple, 2 papers. Single, no dependents: one W-2 and a bank's 1099-INT. The quickest way to a whole return.
  2. The Smith family: a 65-paper extravaganza. Two incomes, a retirement, a college student, a small farm, a rental, stock sales and K-1s, on 37 kinds of tax papers. The latest automatic QA walk took all 65 to an estimate, with nothing left on Review.

To try one:

  1. Download its papers.
  2. Open the App on the Libertas Cloud Hub, on your phone or in your browser.
  3. Tap Import with AI in the Document center and pick up to eight papers. AI works out what each paper is, reads its boxes and files it in the right place.
  4. Check and save. Nothing goes in until you approve it.
  5. Answer what no paper says, and get an estimate where every line explains itself.

AI Import and AI Search are on in the Libertas Cloud Hub, within a monthly usage cap.

Try Libertas →

Search Libertas