itr-wala

File your Indian income tax return from your terminal. No CA, no ₹3,000 fee, no 3 hours on the portal. Every rupee of tax math computed by tested code, not by an LLM.

⏳ AY 2026-27 due dates (ITR-1/2 → 31 July, ITR-3/4 non-audit → 31 August) have passed. You can still file a belated return until 31 December 2026: late fee ₹1,000 if income is up to ₹5 lakh, else ₹5,000 (section 234F), plus interest on any unpaid tax.

Seen in the wild: a reel by @ezsnippet (3.7M followers) walked through this repo: "ab tum bina CA ke bhi Income Tax return bhar sakte ho, ITR-Wala use karke." 1.2M views in a week, and the repo's busiest week so far. Filed with itr-wala? Two lines in this thread help the next filer.

# Review-first (recommended): clone, read it, install the bytes you just read
git clone https://github.com/karanb192/itr-wala.git
cd itr-wala && ./install.sh          # also: codex, gemini, all

# Or scope it to just the folder your tax documents live in
cd ~/tax-2026 && ~/itr-wala/install.sh --here

# Claude Code plugin (name whichever repo you trust)
/plugin marketplace add karanb192/itr-wala
/plugin install itr-wala@itr-wala

# Codex plugin (same repo doubles as a Codex marketplace)
codex plugin marketplace add https://github.com/karanb192/itr-wala
codex plugin add itr-wala@itr-wala

# Or the skills.sh one-liner (fetches the default branch head; pick your agent when prompted)
npx skills add karanb192/itr-wala

Run from a checkout, install.sh never touches the network - it copies the files you just read. See Install from a branch you reviewed for why that matters when the tool handles your salary and bank data.

Then open your agent and say "file my ITR". Hand it your Form 16 and AIS. It does the rest - except the three things only you should ever do: pay, submit, e-verify.

Why this exists

Every AI-tax demo you saw this season had the same silent flaw: the model was doing the arithmetic. LLMs are magnificent at reading a Form 16 and terrible at applying s.87A marginal relief. One transposed digit and your "free filing" costs you a tax notice.

itr-wala splits the work the way it should be split:

The math is defended in three layers, all shipped in the repo and run in CI on every commit:

1. 47 golden tests - every expected value hand-derived from the statute first: the 87A rebate cliff and its marginal relief, capital-gains exemption ordering, s.71 loss set-off, the surcharge tiers (including the exclusive-income tests for the 25%/37% slabs and the 15% ceiling on capital-gains tax), and 234A/B/C/F interest down to the month-counting and challan-date edge cases.
2. A 104-test suite for the input validator - the gate that rejects malformed, mistyped, or PAN-bearing inputs before they can reach the engine.
3. A property-based fuzzer (scripts/fuzz_engine.py) - generates thousands of randomized, boundary-biased returns and asserts invariants the law implies: more income can never mean less tax in the new regime, cess is exactly 4%, rounding follows s.288A/288B, recommendations match the cheaper legal option. Seeded and deterministic; CI replays 3,000 cases on every commit, and 350,000+ were swept before release.

The skill runs the golden suite in front of you before touching your return:

$ python3 skills/itr-wala/scripts/test_tax_engine.py
...............................................
Ran 47 tests in 0.002s
OK

(Installed as a plugin and can't find the path? Just ask the agent to "run the itr-wala self-test".)

If your CA can show you their test suite, hire them.

These layers exist because they catch real bugs. Hand-deriving every scenario caught an early build that denied surcharge marginal relief on capital-gains-heavy incomes, and the fuzzer caught a one-in-350,000 floating-point rounding edge where ₹52,880 more salary computed ₹10 less tax. Both are fixed and pinned as regression tests. That find-fix-pin loop is the thing a prompt-only tax tool cannot run.

What a session looks like

1. Self-test - the engine proves its math before you trust it.
2. Documents - drop Form 16 + AIS (JSON) into a folder; it tells you exactly where to download each one.
3. Extract & validate - every number transcribed verbatim into income.json, then a strict validator cross-checks totals against your documents. Unknown key? Rejected. TDS doesn't match 26AS? Flagged.
4. Deduction hunt - a proactive interview (80C, 80D, NPS, HRA, home loan…), because the portal will never ask you.
5. Both regimes, computed - a comparison table with the exact rupee savings. The regime gap is routinely five figures; this table is where it shows up.
6. Filing pack - every portal field mapped to its value, in order, plus the final payable/refund figure the portal must match to the rupee.
7. The portal, together - it narrates each schedule; you type. It never sees your password or OTP. You alone click Pay, Submit, and e-Verify.

The artifact you actually share

Real output, reproducible from the bundled (fictional) example - python3 skills/itr-wala/scripts/tax_engine.py skills/itr-wala/assets/example-income.json:

Income-tax computation - FY 2025-26 (AY 2026-27)
================================================================
[NEW REGIME]
  Gross total income               26,06,700
  Total income                     25,06,700
  TOTAL TAX                         3,00,350
  NET PAYABLE (-ve=refund)               720

[OLD REGIME]
  Gross total income               22,09,300
  Total income                     18,74,300
  TOTAL TAX                         3,39,730
  NET PAYABLE (-ve=refund)            43,530

================================================================
  RECOMMENDED: NEW regime (saves Rs. 42,811)

Privacy, honestly

• The Python scripts run entirely on your machine. Tax math never leaves.
• Documents you ask the AI to read are processed by the model - that part does leave your machine, like anything you paste into an AI tool. The skill tells you this up front and invites you to redact PAN/Aadhaar/account numbers first: they're not needed for computation, and that's enforced as a mechanism, not a plea - the validator rejects any input file containing a PAN-shaped or Aadhaar-shaped string.
• A generated .gitignore keeps tax documents out of your repos.
• Built by someone who files his own taxes with it - and who happens to do security for a living (DEFCON/OWASP speaker, Head of AI at an application-security company).

What it covers (and refuses)

In scope (AY 2026-27, resident individuals): salary (multiple employers, retirement exemptions like gratuity and leave encashment in both regimes), house property including s.71 loss set-off, equity/MF capital gains (111A/112A/112, grandfathering-aware exemption ordering), debt MF, crypto/VDA, lottery and online-game winnings (115BB/115BBJ), interest & dividends, family pension with the s.57(iia) deduction, s.89 arrears relief, presumptive income (44AD/ADA basics), all Chapter VI-A deductions, both regimes, surcharge with marginal relief, advance-tax interest computed to actual challan dates, late fees, belated returns (including the s.115BAC(6) rule that locks belated filers out of the old regime - it will tell you, not let you find out from a notice), ITR-1/2/3/4 form selection.

Out of scope - it will say so and point you to a CA rather than guess: non-residents/RNOR, F&O and intraday, audit cases, foreign tax credit (Form 67/DTAA), ESOP deferral, the property indexation option, buyback capital-loss entries, agricultural income above ₹5,000. Partial coverage is computed honestly; the rest is never silently approximated.

Hard boundaries, always: never your password or OTP, never clicks Pay/Submit/e-Verify, never fabricates a deduction. Lowest legal tax.

FAQ

Can I trust an LLM with my taxes? No - that's the point. You're trusting a tested Python engine with the math and an LLM with reading PDFs and explaining things, which are the two things each is actually good at. Run the test suite yourself.

But the LLM still reads the documents - what if it misreads a number? True, and worth being precise about: transcription is the one step the model touches, so a misread digit is the residual risk. That's why every figure is cross-checked against independent documents (Form 16 vs 26AS vs AIS - a single-document misread fails validation), recorded next to its source citation, and shown to you in the filing pack before anything is filed. If the model misreads and every cross-check misses it, you'll see the wrong number with its citation - not a hidden one.

Why not just use ClearTax/Quicko/a CA? Use whatever you trust. This is for people who'd rather review every number themselves than pay ₹3,000+ to hope someone else did. The filing pack it generates is also a great ₹0 first draft to hand a CA for a chea