Tax Studio

A private, local-first tax document workspace for VS Code.

Drop your tax forms, expense receipts, and bank statements into an inbox. Tax Studio classifies them, files them into a clean year-first library per taxpayer, extracts the values that matter, and grounds every extracted value to the exact region of the original document — all on your machine. Nothing is uploaded.

Learn more at lynxdi.com/tax.

Tax Studio never computes tax liability and never files returns. It organizes your documents and produces clean, reviewable, source-anchored data — exported as FDX (Financial Data Exchange) JSON — that you or your CPA can trust.

Why it's different

• Local-first, private by default. Your documents stay in your folder. The entire ingest path — classification, extraction, the SQLite evidence database, PDF preview — runs offline. No account, no server, no telemetry. The only optional network touchpoint is a consent-gated AI tier that is off by default (and can be pointed at a fully-local model).
• Deterministic-first extraction. Tax Studio reads structured data straight from the PDF before it ever reasons about text. An AI model is the last resort, not the first.
• Every value is traceable. Each extracted number carries a source anchor — the document, page, region, and form field it came from. Click a value; the original PDF highlights exactly where it lives. No black-box guesses.
• Your files are never altered. Originals are copied/renamed into the library by the normalizer with a post-move hash re-verify; the original bytes are never modified.
• Agent-ready. The workspace is self-describing and ships a read-only MCP server, so any coding agent (Claude, etc.) can safely query your tax data — and propose corrections through one scoped, audited write path.

Tax Studio implements the architecture of US provisional patent 64/115,885, "In-Situ, Model-Independent Document Intelligence with Source-Role-Validated Evidence Grounding, Region-Level Selective Invalidation, and Evidence-Bound Deterministic Generation."

How it works

  input\  ──▶  classify  ──▶  extract (deterministic-first)  ──▶  library\<year>\<taxpayer>\<category>\
 (inbox)        (doc type,       (values + source anchors)          (renamed, deduped, hash-verified)
                 taxpayer,                    │
                 tax year)                    ▼
                                      SQLite evidence graph  ──▶  Review · Analytics · FDX / CPA export

1. Open a folder in VS Code and initialize it — each folder is an independent tax workspace (open a different folder in another window for a different taxpayer).
2. Drop documents into the workspace's input\ inbox (or point Tax Studio at external input folders — originals stay put, copies are filed).
3. Classify & File. Tax Studio detects the document type, the taxpayer, and the tax year, then files it into library\<year>\<taxpayer>\<category>\ with a clean, human-readable name. Taxpayers auto-create from the documents themselves — zero setup.
4. Review extracted values side-by-side with the source PDF; approve, correct, or flag.
5. Export clean FDX JSON, or a CPA review package, when you're ready to hand off.

Features

Ingest & organize

• Zero-setup inbox → library. Documents are classified and filed automatically; taxpayer entities are created from the documents (individuals, Joint, and Business\<Name>\).
• Year-first library — library\<year>\<taxpayer>\<category>\ — with human-readable file names (2025 - 1099-DIV - Example Brokerage - x1234.pdf).
• Returns & planning trees you manage yourself: returns\<year>\<taxpayer>\draft|final\ and planning\<year>\. Filed returns are indexed as prior-year evidence but never renamed.
• Multiple input folders — add external drop zones; originals are kept, copies are filed.
• Exact-duplicate detection, expected-but-missing document checks (vs. the prior year), and year-over-year inventory diffs.
• Two views of the tree: disk-aligned by section, or pivoted by taxpayer — toggle from the title bar.

Deterministic-first extraction pipeline

Tax Studio tries the cheapest, most reliable method first and only escalates when it must. The tier order is fixed:

Bundled form packs cover common forms — W-2, the 1099 family (INT/DIV/B/R/NEC/MISC/OID), 1098/1098-T/1098-E, 5498, 1095-A/B/C, and 1040 / Schedule B — each mapping a form box to a semantic key, an FDX path, and a plain-English description.

Evidence graph with source anchors

Every extracted value becomes an assertion in a local SQLite evidence graph, bound to a source anchor (document · version · page · region · form field). When a source document changes, Tax Studio selectively invalidates only the assertions that depended on the changed region — preserving your human review and corrections everywhere else.

Document Review panel

Open any filed document into a split view: the original PDF on the left (with evidence regions highlighted), its assembled FDX JSON on the right. Click a value to jump to its source region. Values are editable inline — a correction is saved to the database while the machine's original value is preserved as history, and re-extraction conflict-checks instead of overwriting. Fields a form pack expected but that came back empty are flagged in red.

Intake worksheet

A per-year, schema-driven intake worksheet (filing status, dependents, income, deductions, credits, estimated payments, and more) with a sliding Yes/No toggle and auto-fill from last year's answers — a guided way to make sure nothing is missing before you file.

Analytics & planning (sums only — never liability)

Per-taxpayer / per-year rollups of the extracted evidence (W-2 wages, 1099 income by type, business income vs. expense, donations), a document checklist, duplicates, needs-review counts, and a starter CPA-questions doc. Totals are sums of your documents only — Tax Studio never computes tax owed.

Export

• FDX JSON (independently-written, FDX-shaped schemas; targets FDX v6.5.0), per taxpayer and year.
• CPA review package — a foldered handoff of the reviewed data and its provenance.

Model Context Protocol (MCP) — agent access to your tax data

Tax Studio ships a read-only MCP server so a coding agent can query your evidence graph without ever reading the extension's source or touching raw files. Set Up MCP for This Workspace writes an .mcp.json and stages the server; it opens tax.db read-only and answers over stable views. It reads the database through Node's built-in node:sqlite — the native SQLite module is never exposed to the agent surface.

Every tool is read-only except one scoped, audited correction path:

The working folder also generates byte-identical AGENTS.md / CLAUDE.md / GEMINI.md from the live database schema, so any agent can operate on your data read-only without guessing.

Privacy & AI

• No document content ever leaves your machine unless you explicitly turn on the AI tier.
• lynxTax.allowAiRequests is off by default. With it on, you choose the backend:
  • Ollama — a local model, fully offline; or
  • Anthropic — document excerpts are sent to the Anthropic API (requires ANTHROPIC_API_KEY). Default model claude-haiku-4-5.
• The AI tier only runs on documents the deterministic tiers can't handle, and any value it returns must re-anchor to source text or it's confidence-capped and flagged for review.
• Tax Studio runs only in a trusted workspace (it reads your documents, runs a local Python sidecar, and loads a native SQLite module).

Getting started

1. Install the extension and open the folder you want to use as a tax workspace.
2. Run Tax Studio: Initialize Tax Workspace… (or click Initialize This Folder in the Tax Workspace view). This scaffolds input\, library\, returns\, planning\, exports\, and a hidden .tax\ (the database, derived text, and logs).
3. (Optional) Run Load Sample Documents… to fetch a synthetic Jane/John Doe demo bundle (obviously-fake identities) into input\.
4. Drop your own PDFs / images / CSV / OFX / QFX into input\.
5. Run Classify & File Inbox Documents and open the Dashboard.

Key commands (Command Palette → "Tax Studio")

• Initialize Tax Workspace…, Load Sample Documents…
• Scan Inbox, Classify & File Inbox Documents, Process This Document…
• Open Dashboard, Preview Document, Open Tax Worksheet…
• Approve Extracted Value, Flag for Review, Reclassify Document…, Show Evidence in Source
• Find Duplicate Documents, Compare Tax Years…, Check Expected-but-Missing Documents
• Export FDX JSON…, Export CPA Review Package…
• Set Up MCP for This Workspace, Set Up Extraction Backend (Python), Regenerate Working-Folder Agent Map

Requirements

• VS Code ≥ 1.102, Node ≥ 20 (bundled with VS Code's runtime).
• Python ≥ 3.10 (optional but recommended) for deep PDF inspection, table extraction, and OCR. Run Set Up Extraction Backend (Python) to create a managed virtual environment (installs PyMuPDF); OCR additionally uses a system tesseract binary. Without Python, AcroForm and digital-text extraction still work — you just won't get OCR or table tiers.

Settings

Roadmap

Tax Studio's v1 organizes and grounds tax data. The roadmap keeps every extension point that IRS-ready filing will need (form-line mappings, calculation lineage, per-year versioning) so filing can ship in the near future.

• MeF filing readiness — return worksheets → IRS MeF XML generation validated against locally obtained IRS schemas and business rules → transmitter submission. v1 still does not compute liability or file.
• Real IRS form packs & year-versioned packs (in progress) — fill/read the official IRS AcroForm PDFs; select the correct pack for each document's tax year so history years use that year's form.
• Transactions & reconciliation — populate the transaction ledger, OFX/CSV import mapping, and receipt ↔ transaction matching.
• More form packs — K-1 and beyond.
• Broader AI tier rollout and near-text (fuzzy) duplicate detection.
• In-situ mode — index documents in place, without copying them into a library.

The disclosed techniques are domain-general (the patent covers legal, insurance, medical, mortgage, education, compliance, and more); tax is the first shipping embodiment.

Privacy, data, and disclaimer

Tax Studio is a document-organization tool. It does not provide tax, legal, or financial advice, does not compute tax liability, and does not file returns. Always review extracted values against your originals and consult a qualified professional before filing. You are responsible for the accuracy of your return.

About

Tax Studio is developed by Lynx DI. See lynxdi.com/tax.

Built on the architecture of US provisional patent 64/115,885. Third-party product and company names are trademarks of their respective owners, used for identification only.

License: see the LICENSE file.