Field note

Moving an AI workstation from Windows to a Mac

A new Mac, about 80 repos, and one question: which of these projects needs which tool. The answer became an interactive guide.

October 2026 / Setup, short read

01 · The problem

The repos came across. The tools did not.

I moved to a new MacBook after years on Windows. Getting about 80 repositories onto it took one clone loop. Getting them to run took much longer, because a repo does not say which of Node, Python, Rust, Java, Docker, LaTeX, ffmpeg or a speech model it needs. You find out one failed command at a time.

So I stopped installing and started measuring. I wrote a scanner that reads the manifests in the cloned repos and reports what each one needs. It installs nothing. On my machine it found Python in 27 repos, Node in 17 (counting the pnpm ones) and Docker in 6, and it named the rest. I installed what the repos I use needed, one step at a time, with a check after each.

02 · What was not obvious

Three things I would have guessed wrong

A coding agent cannot type your password, so the setup has to say which steps are yours. Homebrew casks such as BasicTeX run an installer with sudo, and an agent's shell has no terminal to ask. Those steps stay with you, and the plan should mark them.
Guard every write. At one point my agent overwrote an AGENTS.md that Codex had written moments earlier, because it used a plain cat >. I recovered the text from Codex's session log. Every file the guide writes now starts with a check that the file does not already exist.
"Needs CUDA" often means "written for NVIDIA". A 0.6B voice-cloning model ran at 0.66 to 0.82 times real time on the Mac's GPU. On the CPU, a six-sentence test had not finished after 28 minutes. The repo hardcoded cuda:0. Adding an mps option took about twenty lines.

The pattern every write in the guide follows

[ -f ~/.config/ghostty/config ] || cat > ~/.config/ghostty/config <<'EOF'
copy-on-select = clipboard
EOF

Try it

The guide is at tech.anujsadani.in/mac-ai-dev-setup. Tick what your projects need and copy the plan, or paste your own scanner output to tick the boxes for you. Each step carries a label saying how it was verified: run on the Mac, tested in a scratch home folder, or a standard command that was not run here. The scripts and tests are in the site repository.

03 · Limits

One machine, one measurement

This is a short note from a single move. The Mac is an M5 Pro with 24 GB of memory, and the speeds above are one measurement on it, not a benchmark. The Homebrew, Claude Code and Docker installs in the guide are standard commands I did not re-run for it, and the guide says so on those steps.