atharva262005/ai-audit-shelf

ai-audit-shelf is a Python tool that adds Git-like version tracking to AI workflows, so every prompt and result an AI system produces gets recorded and can be traced later. It organizes this tracking around a library metaphor: individual AI actions become permanent chapters, chapters are bundled into versioned books that represent a feature, and books sit together on a shelf grouped by feature. The problem the author is addressing is that AI workflows are often opaque. Prompts go in, results come out, and there is usually no record of who ran what, when, or why. This tool logs every prompt, its result, the actor who triggered it, and a timestamp, and keeps that record immutable. When a feature is edited, a new edition of its book is created rather than overwriting the old one, so the original stays available as a permanent audit trail. Records can be exported as JSON for machines or as Markdown for human auditors. There are three ways to use it. A command line tool lets someone log chapters, bundle them into books, browse the shelf, compare two editions, export records, and search by actor, keyword, or date, using only Python's standard library. A REST API server, built with FastAPI and Uvicorn, exposes the same operations over HTTP and can optionally require an API key on write requests. A web dashboard connects to that API and gives a visual way to browse the library, view books in a table with export buttons, inspect the full chapter log, search, and diff two editions side by side. The README also documents integration recipes for plugging the audit log into real AI systems, including a plain Python requests example, a LangChain callback handler that logs every LLM call and tool call automatically, and an example using OpenAI function calls. Together these make it possible to bolt an audit trail onto an existing agent or chatbot with only a few lines of code, without changing how the underlying AI system works.