Hi Product Hunt! 👋 I'm Jet, the maker of DocMason. As an IT Architect, I swim in complex office documents every day — decks, spreadsheets, legal PDFs, architecture diagrams from every team. Normal LLM chat windows completely fail here: they flatten everything, lose structure, and forget context the moment you close the tab. So I started building DocMason a few weeks ago to solve my own problem. I now use it every single day. Then Andrej Karpathy posted about "LLM Knowledge Bases" — and it perfectly validated the direction. Here's the thing: most ChatGPT Plus users are sitting on powerful, unused Codex capacity simply because they don't write code. DocMason turns that idle capacity into something genuinely useful for office work. The paradigm I'm exploring: The repo is the app. Codex is the runtime. DocMason takes Karpathy's concept further for real-world use: • Handles messy office files — .pptx, .docx, .xlsx, .eml — and truly extracts multimodal content like IT architecture diagrams and complex spreadsheets • Runs as a real app, not naive RAG — auto-prepares environment, auto-syncs knowledge base incrementally • Fully agent-native — runs inside Codex or Claude Code, giving your documents the same autonomous treatment engineers give codebases It's fully open-source and runs locally (your files never leave your machine). Check the demo video & architecture diagram in the GitHub Readme and give it a try! What type of office file breaks your current AI workflow the most? 👇

DocMason

Agentic app for deep research over your private office files

Agentic app for deep research over your private office files