The problem
Feeding a codebase to an LLM usually means copy-pasting files by hand or running a black-box tool that hides what it actually includes.
What I built
llmScribe maps a project, searches it, and retrieves exactly what you need — no AI, no embeddings, no hidden ranking. It is dependency- and git-aware, works on local folders or public GitHub repos without cloning, and ships as a command-line tool, a desktop app and a hosted server for AI agents.
How I built it
Pure Python, published on PyPI under Apache-2.0. One core powers three interfaces: a CLI, a Tkinter GUI, and a hosted MCP server for AI agents.
Key decisions
- Kept the output deterministic and transparent — plain structured text, so you can see exactly what goes to the model.
- Shipped three interfaces — command line, desktop and server — from one core so it fits both humans and agents.
Result
Open-source and published on PyPI under Apache-2.0, with stars on GitHub and real use as an MCP server. It is the project that best shows the Python-plus-system-design side of how I work.