CBCharlie Barmore
← All projects
TaxClaw logo

TaxClaw

The tax half of LedgerClaw: AI reads the documents, a real engine computes the return, and it ties out on a canvas workpaper, with the CPA in control before anything is filed.

In development For solo & small CPA firms

All product screens on this page show demo data, never real client or user information. They’re here to demonstrate what the product can do.

TaxClaw compliance workflow for a 2025 return: a seven-phase pipeline (Intake → Workpaper → Projection → Client Review → Sign & File → Filed → Closed), an active blocker for a missing K-1, an expected-documents checklist auto-populated by source, and a workpaper where dropped W-2s, 1099s, and 1098s are read, extracted, and tied out, with a one-click Export for Lacerte.

About

TaxClaw is the tax half of the workflow. Where LedgerClaw keeps the books all year, TaxClaw creates the tax workpapers and works up the return. LedgerClaw has only light tax functionality, so TaxClaw is the real thing: turning a shoebox of client documents, and a year of clean books, into a computed, tied-out return, with the reading done by AI, the math done deterministically, and the CPA in control of everything before it’s filed. It runs in my own practice today, and it’s on the same path to availability LedgerClaw is already on.

The flow mirrors how a preparer actually works:

  • Import the documents. Drop in W-2s, 1099s, K-1s, and 1098s; TaxClaw reads the boxes, files each as structured data, and tracks what’s still outstanding.
  • Build the workpaper. Extracted forms roll up by category and reconcile against last year; missing, new, and changed items get flagged.
  • Compute the return. A deterministic Form 1040 + Georgia engine calculates the return line by line (not an AI guess) and explains every number.
  • Tie it out on a canvas-style workpaper. Toggle from the standard forms into an infinite-canvas workpaper: drag documents into sections, watch subtotals add up live, and crossfoot the whole thing back to the 1040 in real time, with notes, a calculator tape, and a link-arrow audit trail.
  • Review, plan, and sign off. Flag items for review, run estimates and planning, generate client deliverables, and export to the filing software. Nothing is final until the CPA signs off.

It automates the grunt work (reading forms, chasing documents, tying out numbers) without handing over control of the return.

TaxClaw vs. LedgerClaw: LedgerClaw keeps the books right all year; TaxClaw creates the tax workpapers and works up the return. Same DNA: AI reads documents into structured data, the CPA reviews. Just a different deliverable and cadence.

The canvas began as a separate prototype, CanvasClaw, now merged in as TaxClaw’s spatial workpaper view. It’s private for now, running in my own practice, and on the same path to availability LedgerClaw is already on, once it’s hardened for client tax data and the engines are complete.

How I built it

The core idea: AI does the reading, a real engine does the math, and the CPA stays in the loop.

  • A real tax engine, not a prompt. The math is a hand-built, deterministic Form 1040 + Georgia engine with “rules as data”: each year’s brackets and limits are swappable parameters. It covers the hard parts: QBI / §199A (with the wage and SSTB limits), AMT, NIIT, taxable Social Security, Schedule D, and CTC/ACTC.
  • Proved correct, independently. The engine is validated to the dollar against PolicyEngine, a separate open-source tax model, with answers fetched live at test time so the check can’t be circular. 24 federal and 7 Georgia scenarios match exactly.
  • Safe by default. Documents are read by Claude; numbers are computed deterministically; every action is logged and reviewable. Anything the engine can’t model is flagged for review, never silently guessed.
  • Two apps merged into one. The canvas was a separate Vite + Electron prototype that I ported into TaxClaw’s Next.js stack as a view layer over the existing engine, reusing the backend instead of duplicating it.
  • Shipped end to end. Local-first prototype → hosted production: Next.js, Postgres, Clerk auth, and field-level encryption.
  • White-labeled by design. Today TaxClaw wears my firm’s color palette and CB monogram, since it runs inside my practice. The same theming is built to white-label to any firm down the road.

Stack

Next.jsTypeScriptPostgresClerkClaude Code
The client cockpit: outstanding docs, open blockers, and tasks up top; the active return with its engine-computed projection; and a live feed of agent runs and notes down the right. The engine flagged a $3,747 balance due straight off the documents.
The client cockpit: outstanding docs, open blockers, and tasks up top; the active return with its engine-computed projection; and a live feed of agent runs and notes down the right. The engine flagged a $3,747 balance due straight off the documents.
Every fact, traced to its source: AI reads the organizer and source documents into structured inputs (filing status, dependents, AGI, Schedule C, deduction method), each one citing the document it came from. Manual overrides win over re-extraction.
Every fact, traced to its source: AI reads the organizer and source documents into structured inputs (filing status, dependents, AGI, Schedule C, deduction method), each one citing the document it came from. Manual overrides win over re-extraction.
The canvas workpaper: one click flips from standard forms to an infinite canvas where Wages, Investment, Retirement, and Property sections subtotal their documents live, every extracted W-2, 1099, and 1098 carries its box-level data with a view-source link back to the original document, and Crossfoot ties the sheet back to the return.
The canvas workpaper: one click flips from standard forms to an infinite canvas where Wages, Investment, Retirement, and Property sections subtotal their documents live, every extracted W-2, 1099, and 1098 carries its box-level data with a view-source link back to the original document, and Crossfoot ties the sheet back to the return.
One agent per workflow phase, intake through pre-filing QC: Docs & Intake, Preparer, Return Review, Tax Planning, and Quality Control. Each run is logged with its duration and token cost, and anything an agent proposes waits for the CPA's approval before it executes.
One agent per workflow phase, intake through pre-filing QC: Docs & Intake, Preparer, Return Review, Tax Planning, and Quality Control. Each run is logged with its duration and token cost, and anything an agent proposes waits for the CPA's approval before it executes.