My Claude Code stack

Claude Code 2.1.296

Anthropic's agentic coding CLI. Runs on macOS, Linux and Windows. I run it from my home folder and point it at project folders as needed. I usually work in auto mode, where Claude acts without asking at every step.

Models Opus 5.5

• Opus 5.5: my default, effort high
• Sonnet 5: effort xhigh when I switch to it
• /model switches models; /fast turns on fast mode

Terminal status & notifications

Every Claude Code hook event runs iTerm2's cc-status helper, so each tab shows whether Claude is working or waiting for me.

On my Mac, hooks feed iTerm2 so each tab shows whether Claude is working or waiting. iTerm2 is Mac-only. The equivalent here is a hook that pops a desktop notification when Claude finishes or needs you.

I use iTerm2. Check whether ~/.config/iterm2/cc-status exists. If it does, add a hook to ~/.claude/settings.json that runs it on every event (SessionStart, SessionEnd, UserPromptSubmit, PreToolUse, PostToolUse, PermissionRequest, Notification, Stop, StopFailure, SubagentStop), and set preferredNotifChannel to "iterm2".

Add hooks to ~/.claude/settings.json so I get a desktop notification (notify-send, install libnotify-bin if missing) when you need my input (Notification and PermissionRequest events) and when you finish a task (Stop event). Test that a notification appears.

Add hooks to my Claude Code settings.json so I get a Windows toast notification when you need my input (Notification and PermissionRequest events) and when you finish a task (Stop event). Use PowerShell with the BurntToast module (install it for my user if missing). Test that a toast appears.

Persistent memory

Claude Code saves what it learns about my projects and preferences to small notes in its .claude/projects/…/memory/ folder, with a MEMORY.md index that loads at the start of every session. It's built in on every OS. Tell Claude to “remember” things and it will.

From now on, when you learn something non-obvious about how one of my projects builds, deploys or tests, save it to memory with the date. Keep one fact per memory file.

Universal Modder rehan-remade

Skills and a um CLI for modding PC games, plus a sprite/asset pipeline (um sprite fit, um sprite sheet, um comfy run). I mostly use it for the asset tools with my local ComfyUI. Its fal MCP server needs your own fal.ai key (pay-as-you-go). Skip that part if you don't want it. It needs uv and ImageMagick (see Toolchain).

/plugin marketplace add rehan-remade/universal-modder
/plugin install universal-modder@universal-modder

Language servers official

From Anthropic's official marketplace. They give Claude real go-to-definition and diagnostics in code.

• clangd-lsp: C/C++, needs clangd
• swift-lsp: Swift, needs sourcekit-lsp from a Swift toolchain

xcode-select --install

sudo apt install -y clangd
# Swift (optional): see swift.org/install/linux

winget install -e --id LLVM.LLVM
# Swift (optional):
winget install -e --id Swift.Toolchain

/plugin install clangd-lsp@claude-plugins-official
/plugin install swift-lsp@claude-plugins-official

Document skills

• docx, pptx, xlsx, pdf: read and write Office files and PDFs
• docs: Claude Docs, shareable docs people can comment on
• google-workspace: edit Google Docs, Sheets and Slides

Research & prompting

• deep-research: multi-source research reports
• prompt-master: writes prompts for other AI tools
• skill-creator: build and test your own skills
• import-memory: bring memories over from ChatGPT etc.

Browser & computer control

• chrome-browser: drives your real Chrome through the Claude in Chrome extension (works on all three OSes)
• computer-use / built-in-browser: need the Claude desktop app (macOS and Windows only, so not available on Linux)

Artifacts

Claude Code can publish HTML pages, dashboards and decks to a private claude.ai link that you can choose to share. A page like this one could be published that way.

Everyday CLI

• git, ripgrep, jq
• uv: fast Python package and venv manager
• tmux: terminal multiplexer (WSL only on Windows)

brew install ripgrep jq tmux uv bash mawk

sudo apt install -y git ripgrep jq tmux
curl -LsSf https://astral.sh/uv/install.sh | sh

scoop install git ripgrep jq uv
# tmux: install WSL, then `sudo apt install tmux` inside it
wsl --install

Web / Drupal sites

• Node 24 and Node 12 via nvmnvm-windows (old gulp builds need 12)
• PHP 8 + Composer
• Gulp + Sass theme builds; legacy Sass 3.2 / CoffeeScript 1.7 for Drupal 7 themes
• Headless Chrome for screenshots and before/after checks
• Database GUI: TablePlusTablePlusDBeaver (TablePlus's Linux build is limited)
• SFTP: TransmitWinSCP (Transmit is Mac-only)FileZilla (Transmit is Mac-only)

nvm install 12 && nvm install 24
brew install php composer
brew install --cask tableplus transmit

nvm install 12 && nvm install 24
sudo apt install -y php-cli php-xml php-mbstring php-gd php-curl php-zip composer
sudo snap install dbeaver-ce
sudo apt install -y filezilla

nvm install 12
nvm install 24
scoop install php composer
winget install -e --id WinSCP.WinSCP
# TablePlus: download from tableplus.com

Apple / iOS

• Xcode, Swift 6.4, xcodebuild, codesign
• libimobiledevice (idevice_id, afcclient) to talk to an iPhone over USB

Building iOS/macOS apps needs Xcode, which only runs on a Mac. The Swift language itself runs here (see Language servers above). libimobiledevice works on Linux too. libimobiledevice has no good native Windows build, so skip it.

xcode-select --install
brew install libimobiledevice

sudo apt install -y libimobiledevice-utils usbmuxd

Game dev

• Godot 4.7, with a godot command on the PATH so Claude can run headless tests
• devkitPro (devkitARM, libctru, citro2d) for Nintendo 3DS homebrew
• Azahar 3DS emulator for testing

brew install --cask godot
mkdir -p ~/.local/bin
ln -s /Applications/Godot.app/Contents/MacOS/Godot ~/.local/bin/godot

# Godot: download the Linux .zip from godotengine.org, then:
mkdir -p ~/.local/bin ~/apps
unzip Godot_v4*_linux.x86_64.zip -d ~/apps
ln -s ~/apps/Godot_v4*_linux.x86_64 ~/.local/bin/godot

# devkitPro (Debian/Ubuntu):
wget https://apt.devkitpro.org/install-devkitpro-pacman
chmod +x ./install-devkitpro-pacman
sudo ./install-devkitpro-pacman
sudo dkp-pacman -S 3ds-dev

winget install -e --id GodotEngine.GodotEngine
# devkitPro: run the graphical installer from
# github.com/devkitPro/installer/releases and tick "3DS Development"

For the headless godot command, put the Godot_*_console.exe from the godotengine.org .zip on your PATH as godot.exe.

Images & assets

• ImageMagick: resize, crop, alpha checks
• oxipng + pngquant: PNG compression
• webp: WebP conversion
• PhotoshopGIMP or Krita (Photoshop doesn't run on Linux) for manual touch-ups

brew install imagemagick oxipng pngquant webp

sudo apt install -y imagemagick pngquant webp
sudo apt install -y oxipng   # Ubuntu 24.04+ / Debian 12+; else see GitHub releases
sudo apt install -y gimp krita

scoop install imagemagick oxipng pngquant libwebp

Ubuntu's apt ImageMagick is often version 6, which uses convert instead of magick. Claude handles either.

Prompt: install a group for me

Check which of these are already installed on my Mac and install only what's missing with Homebrew: ripgrep, jq, tmux, uv, imagemagick, oxipng, pngquant, webp, php, composer, libimobiledevice. Then show me a table of what was installed and the versions.

Detect my Linux distro and package manager. Check which of these are already installed and install only what's missing: ripgrep, jq, tmux, uv, imagemagick, oxipng, pngquant, webp, php (cli + xml/mbstring/gd/curl/zip), composer. If a package isn't in my distro's repos, use the project's official install method. Show me each command before running anything with sudo, then give me a table of what was installed and the versions.

I'm on Windows. Check which of these are already installed and install only what's missing, using winget for apps and Scoop for CLI tools (install Scoop first if needed): ripgrep, jq, uv, imagemagick, oxipng, pngquant, libwebp, php, composer. Then show me a table of what was installed and the versions.

ComfyUI image gen

A node-based image generation server. It runs as a background service bound to localhost. Claude submits saved workflow JSON to it, so it can generate game sprites and site art without a paid API.

• Z-Image Turbo: fast text-to-image (Apache 2.0)
• Qwen-Image-Edit 2511: same character in new poses, views and recolours
• BiRefNet: background removal for transparent sprites
• ComfyUI-GGUF node: runs GGUF models on Apple Silicon (fp8 doesn't work on MPS)

Easiest route on Windows: the ComfyUI Desktop installer, or the portable NVIDIA .7z from the GitHub releases.

Ollama local LLM

Runs open-weight language models locally. Planned: qwen3.6:27b (Q4, about 17 GB) with context around 64K. Size the model to your RAM/VRAM. Claude Code can point at a compatible local endpoint through ANTHROPIC_BASE_URL.

brew install ollama
ollama pull qwen3.6:27b

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3.6:27b

winget install -e --id Ollama.Ollama
ollama pull qwen3.6:27b

Live-site preview proxy

preview-proxy.mjs proxies a live website to localhost and swaps in my locally built CSS/JS, so I can see a change on the real site before deploying. It shuts down after 30 idle minutes. It's plain Node, so it runs on any OS. Only the “open Chrome” helper differs.

Write ~/site-tools/preview-proxy.mjs (plain Node, no dependencies): usage `node preview-proxy.mjs https://site.com 8080 name=/local/file.css ...`. Proxy every request to the live site, serve the local file for any URL containing one of the names, rewrite links back to localhost, strip CSP and HSTS headers, and exit after 30 minutes with no requests. Also write a helper script for my OS that opens URLs in a separate Chrome profile at a width and height I pass in.

Remote ComfyUI helpers

A script that sends a Wake-on-LAN packet to the AI machine and opens an SSH tunnel to its ComfyUI, plus one that uploads a reference image and runs the pose/recolour edit workflow.

I have ComfyUI on another machine at <IP> (user <name>). Write ~/ComfyWorkflows/studio-comfy.sh: send Wake-on-LAN to its MAC address (install wakeonlan if needed), wait for SSH, and open a background tunnel from local port 8188 to its 127.0.0.1:8188. `studio-comfy.sh stop` closes the tunnel. Set up key-based SSH first if it isn't already.

I have ComfyUI on another machine at <IP> (user <name>). Write a PowerShell script %USERPROFILE%\ComfyWorkflows\studio-comfy.ps1: send a Wake-on-LAN magic packet to its MAC address, wait until SSH answers, and open a background tunnel (Windows' built-in OpenSSH) from local port 8188 to its 127.0.0.1:8188. `studio-comfy.ps1 stop` closes the tunnel. Set up key-based SSH first if it isn't already.