TaxHacker (vas3k/TaxHacker) is an open-source self-hosted AI accounting tool aimed at freelancers, indie founders, and small businesses who want LLM-powered transaction categorization without handing financial data to a SaaS vendor. The agent ingests receipts (image, PDF), invoices (PDF, email forwards), and bank statement CSVs, runs OCR-plus-LLM extraction to pull amounts, dates, vendors, and tax categories, and produces structured exports for accountants or tax filing software. Built by Vasily Zubarev (vas3k), the popular Russian-language tech blogger, it ships as a single Docker container and supports local models via Ollama or remote APIs (Claude, GPT-4). Hit 5,490 stars with +106 in the last 24h on May 4, 2026 — a smaller but high-signal entry in the vertical-agent wave, representing the privacy-first SMB accounting niche.
TaxHacker is the privacy-first accounting agent for people who don’t want their financial data going to a SaaS vendor. Self-hosted, single Docker container, Ollama-compatible — point it at a folder of receipts and a bank statement export, and it categorizes everything with LLM-powered extraction, producing tax-ready output for an accountant or filing software.
Built by Vasily Zubarev (vas3k), the agent represents a specific niche in the May 2026 vertical-agent wave: SMB accounting workflows that need LLM intelligence but can’t tolerate data leaving the user’s infrastructure. Hit 5,490 stars on May 4, 2026 with +106 in the last 24 hours — modest velocity, but high signal because the audience is narrow and motivated.
Receipt ingestion: Drop image or PDF receipts into a watched folder. OCR pulls text, then an LLM extraction pass identifies vendor, amount, date, currency, and tax category.
Invoice handling: PDF and email-forward invoices are parsed similarly. Recurring vendors are recognized and categorized consistently.
Bank statement reconciliation: CSV import from major banks; transactions are auto-categorized and matched against ingested receipts/invoices.
Local-first by default: Works fully offline with Ollama-served local models (Llama, Qwen, etc.). Switch to Claude or GPT-4 when extraction accuracy matters more than privacy.
Tax-ready exports: Categorized data exports to formats accountants accept (CSV, structured JSON, jurisdiction-specific templates).
Freelancers and indie founders: Replace bookkeeping spreadsheets and the manual “screenshot every receipt at year-end” workflow with an automated, privacy-respecting agent.
Small businesses with sensitive data: Industries (legal, healthcare-adjacent, regulated) where SaaS accounting platforms are off the table.
International users: Multi-currency and multi-jurisdiction support makes it usable outside US-centric accounting SaaS.
Privacy-conscious users: Anyone who specifically wants their financial data to stay on their own infrastructure.
Self-hosting means you own the operations — backups, updates, and OCR/LLM compute are your responsibility. Local models are accurate enough for most receipts but may lag behind frontier APIs on edge cases (handwritten receipts, foreign-language vendors). Tax categorization is a starting point — final returns should still be reviewed by a qualified accountant, especially for non-trivial tax situations.
Opinionated bundle of agent personas — pre-built role templates for engineering, marketing, ops, and research teams that ship as drop-in markdown files.
OpenAI's flagship AI assistant capable of conversation, analysis, coding, writing, and multimodal reasoning across text and images.
Claude Code session memory plugin that captures context, compresses it via Agent SDK, and injects relevant memory into future sessions automatically.