nheer dotfiles

Install chezmoi with:

sh -c "$(curl -fsLS get.chezmoi.io)" -- -b $HOME/bin

Note

You'll find the repo cloned to ~/.local/share/chezmoi

Install and apply configuration with:

export GITHUB_USERNAME=niklas-heer
$HOME/bin/chezmoi init --apply --ssh $GITHUB_USERNAME

Note

The scripts are run in alphabetical order.

Requirements

• MacOS 15 (Sequoia) or higher
• 1Password and SSH-Agent setup

Used technology

• chezmoi - keeps the dotfiles in sync and secure
• homebrew - installs mac applications
• zsh - the macOS-native shell, enhanced with real-time completions

Local LLM (optional per machine)

The base dotfiles setup does not install an LLM runtime, start a service, or download model weights. On a capable Apple-silicon Mac, opt in explicitly:

setup-local-llm             # desktop apps plus 9B and 27B Ollama profiles
setup-local-llm --fast-only # desktop apps plus only the lower-memory 9B model
setup-local-llm --skip-models # install the apps without model downloads

The command checks the Mac's architecture, OS version, unified memory, and free disk space. It installs Ollama Desktop and oMLX as Mac apps, plus OpenCode as the coding harness. It also migrates an old Homebrew Ollama service to the desktop app without deleting ~/.ollama or its models.

After opting in:

llm                         # Qwen 3.5 9B: fast daily chat
llm --quality               # Qwen 3.6 27B: best local quality
llm "Explain this error"    # one-shot prompt
llm --status                # show loaded model and GPU placement
llm --stop                  # release unified memory
local-ai                    # coding agent using the fast local model
local-ai --quality          # coding agent using the stronger 27B model
local-llm-ui                # open the oMLX app and its Admin UI
local-llm-ui ollama         # open Ollama Desktop

The Ollama profiles use its native MLX engine and a 32K context window. Its API remains on http://127.0.0.1:11434, and cloud features are disabled. oMLX is a separate runtime with a local Admin UI at http://127.0.0.1:8000/admin; its models are selected interactively and are not duplicated automatically. Avoid loading large models in both runtimes at once on a unified-memory Mac.

local-ai runs the Ollama profiles through OpenCode so it has repository context, tools, and an edit/test loop instead of operating as raw chat. Extended thinking is disabled in OpenCode to keep agent loops responsive.

Post tasks

There are some tasks after the install as they cannot be automated.

• Install SetApp apps: CleanShot X
• Download Premium voices on macOS: "Anna (Premium)", "Jamie (Premium)"
  • https://support.apple.com/guide/mac-help/change-the-voice-your-mac-uses-to-speak-text-mchlp2290/mac

Project decision log

I want to document my decisions for me so I don't forget and potentially for you so you understand why I use one tool or another or why I remove stuff from time to time.

The records live in decisions/ as vrdx decision records: one Markdown file each, with TOML metadata between +++ lines.

vrdx list
vrdx show <id-or-prefix>
vrdx new "<title>" --body-file <file>

They are not deployed to the home directory.