Architecture diagrams are almost always out of date. Drawing them by hand is tedious, and AI tools that generate them from code usually want your repository uploaded to a hosted service.
Chalkline is an MIT-licensed desktop app I built to fix that. You can describe a system in plain English, point it at a local code folder, or draw by hand with over 3,800 AWS, Azure, and Google Cloud icons. You then refine the result visually, through AI edits that keep your layout intact, or directly in JSON, and export it as PNG, SVG, or HTML.
Instead of asking for another API key, Chalkline can use the Claude Code, Codex CLI you're already logged into, a local model, or any OpenAI-compatible endpoint. Diagrams stay on your machine, and the code scanner respects .gitignore and redacts common secrets.
In this talk, I'll demo Chalkline live, including having it diagram its own codebase, and cover:
Why I built it when Eraser, GitDiagram, and others already exist.
How it's built: Electron, React, ELK, and Eraser's open-source diagram packages.
What I learned getting an LLM to produce diagrams that are actually correct.
The roadmap and where I'd love help: testing on Linux and Windows, and a possible Tauri port.