wtf is prompt-mac?

prompt-mac is a single command script that turns a fresh Apple Silicon Mac into a development setup built specifically around working with AI coding agents. Created by Kevin Dunglas, the author of several open-source developer tools, it installs a curated set of CLI tools, apps, and editor configurations rather than a generic all-purpose dotfiles setup. Running the script is as simple as pasting one curl command into the stock terminal. It is idempotent, meaning you can run it again later to pick up new changes to its package list or configuration files, and anything already installed is skipped safely. You can also clone the repository and run the install script by hand if you would rather review the code first, though the configuration files are symlinked back into the cloned folder, so that folder needs to stay in place afterward. The philosophy behind the project favors working with macOS defaults instead of overriding them (Safari, Finder, and Spotlight are left untouched), staying minimal rather than shipping thousands of lines of shell configuration, and treating AI coding agents as a first-class part of the setup so their token usage stays under control. The same set of modern command-line tools is meant to feel identical whether you are in a standalone terminal or inside an editor's integrated terminal. Out of the box it configures the Ghostty terminal with GPU acceleration and a quake-style drop-down window, the cmux terminal manager for running multiple AI agents side by side, the starship prompt, and a zsh setup with fast plugin loading. Familiar commands are quietly swapped for faster modern equivalents: ls becomes eza, cat becomes bat, find becomes fd, grep becomes ripgrep, cd becomes zoxide, and vi becomes neovim. Vim-style modal editing is wired into the shell itself, into neovim as the default editor, and into VS Code through a pre-installed Vim extension. On the AI side, the setup includes a tool the README calls a token killer, which compresses shell command output by 60 to 90 percent before it reaches the Claude Code agent, alongside the official Claude desktop app, aimed at keeping API costs and context windows manageable when running heavy agent workloads.