Self-hosted receipt scanner where you write the LLM prompts

What it does

TaxHacker is a Next.js application that ingests photos and PDFs of receipts, invoices, and bank statements, then uses an LLM to extract structured accounting data—dates, amounts, merchants, line items, taxes—into a PostgreSQL database. You can define custom fields and categories, split multi-item invoices into separate transactions, and filter or export the results to CSV. It also converts foreign currencies and fourteen cryptocurrencies to your base rate using historical exchange data from the transaction date.

The interesting bit

Most receipt scanners give you a black-box API and rigid fields; TaxHacker exposes the system prompts and lets you write per-field extraction rules, so you can teach the model to pull project codes, addresses, or whatever your accountant actually needs. It speaks to any OpenAI-compatible endpoint, including local stacks like Ollama or vLLM, which means your financial documents never have to leave your server.

Key highlights

• Custom AI prompts for every field, category, and project, including the base system template
• Local LLM support via OpenAI-compatible APIs (Ollama, LM Studio, vLLM, LocalAI)
• Automatic currency conversion across 170+ fiat currencies and 14 cryptocurrencies using historical rates
• Multi-project organization with full-text search, bulk operations, and CSV export
• Docker-native self-hosting with PostgreSQL; no external SaaS required

Caveats

• The project is explicitly in early development and carries a “use at your own risk” warning
• Local LLM OCR quality depends entirely on the model you choose; the README notes results are “not guaranteed”

Verdict

Freelancers, indie hackers, and small-business developers who currently dump receipts into a shoebox will get the most value—especially if they already run a home server. If you need battle-tested, audit-grade accounting software or hate writing prompts, wait for a stable release or stick to your spreadsheet.