Opinionated lean setup for Claude Code — cross-session memory, token compression, a persistent knowledge brain, and native Excel/Word skills. Plus a library of pre-built skills and MCP server configs.
skills/— installable skills (slash-command surfaces, runnable docs)mcps/— MCP server implementations and configssetup/— per-OS setup prompts you can paste into Claude Code
License: MIT.
Two paths per OS:
- Direct — copy a shell block, paste in a terminal, runs the install end-to-end.
- Agent-driven — paste the corresponding
setup/<os>.mdinto Claude Code as your first message; it walks you through the phases interactively (recommended if you've never set this up).
The shell blocks below trigger GitHub's "copy" button when you hover over them on github.com — one click to clipboard.
Open Terminal (⌘+Space → "Terminal"), paste, run once:
# 1. Homebrew
[ -x "$(command -v brew)" ] || /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
# 2. Core tools
brew install bun node@22 python@3.12 uv jq rtk
# 3. Claude Code
curl -fsSL https://claude.ai/install.sh | sh
# 4. RTK token-savings hook
rtk init -g
echo "Setup complete. Run: claude — then paste setup/macos.md as your first message."→ Full agent-driven prompt: setup/macos.md (paste into Claude Code)
# 1. System packages
if command -v apt >/dev/null; then sudo apt update && sudo apt install -y curl git jq build-essential python3 python3-pip
elif command -v dnf >/dev/null; then sudo dnf install -y curl git jq @development-tools python3 python3-pip
elif command -v pacman >/dev/null; then sudo pacman -S --noconfirm curl git jq base-devel python python-pip
fi
# 2. Bun
[ -x "$(command -v bun)" ] || curl -fsSL https://bun.sh/install | bash
export PATH="$HOME/.bun/bin:$HOME/.local/bin:$PATH"
# 3. uv (Python venv manager)
[ -x "$(command -v uv)" ] || curl -LsSf https://astral.sh/uv/install.sh | sh
# 4. RTK
[ -x "$(command -v rtk)" ] || (curl -fsSL https://github.com/rtk-ai/rtk/releases/latest/download/rtk-x86_64-unknown-linux-gnu.tar.gz | tar xz -C /tmp && sudo install /tmp/rtk /usr/local/bin/rtk)
# 5. Claude Code
curl -fsSL https://claude.ai/install.sh | sh
rtk init -g
echo "Setup complete. Run: claude — then paste setup/linux.md as your first message."→ Full agent-driven prompt: setup/linux.md
Open PowerShell. First-time only:
Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSignedThen:
# 1. Core tools via winget
winget install --id OpenJS.NodeJS.LTS --silent --accept-source-agreements --accept-package-agreements
winget install --id Python.Python.3.12 --silent --accept-source-agreements --accept-package-agreements
winget install --id stedolan.jq --silent --accept-source-agreements --accept-package-agreements
winget install --id Git.Git --silent --accept-source-agreements --accept-package-agreements
$env:Path = [System.Environment]::GetEnvironmentVariable("Path","Machine") + ";" + [System.Environment]::GetEnvironmentVariable("Path","User")
# 2. Bun + uv
if (-not (Get-Command bun -ErrorAction SilentlyContinue)) { irm bun.sh/install.ps1 | iex }
if (-not (Get-Command uv -ErrorAction SilentlyContinue)) { irm https://astral.sh/uv/install.ps1 | iex }
# 3. Claude Code via npm
npm install -g @anthropic-ai/claude-code
# 4. RTK
if (-not (Get-Command rtk -ErrorAction SilentlyContinue)) {
Invoke-WebRequest -Uri "https://github.com/rtk-ai/rtk/releases/latest/download/rtk-x86_64-pc-windows-msvc.zip" -OutFile "$env:TEMP\rtk.zip"
Expand-Archive -Path "$env:TEMP\rtk.zip" -DestinationPath "$env:USERPROFILE\.local\bin" -Force
}
rtk init -g
Write-Host "Setup complete. Run: claude — then paste setup/windows.md as your first message."→ Full agent-driven prompt: setup/windows.md
(WSL2 user? Use the Linux block inside your WSL distro instead — cleaner.)
If you want to run 2+ Codex/Claude agents in parallel on the same repo without them stomping on each other, layer this tier on top of the lean stack:
- recodeee/gitguardex —
gxCLI. Per-agent branch + worktree isolation, file locks, PR-only merges. - recodeee/colony — Local-first MCP for fleet coordination. Replaces 30k-token repo handoffs with ~400-token compact state in SQLite at
~/.colony. Auto-detects file claims, task graphs, and prior decisions so agents see ownership before editing.
→ Full setup prompt: setup/parallel-agents.md (Linux + macOS; Windows via WSL2)
Skip this tier if you only ever run one Claude Code window at a time — claude-mem + gbrain are enough for solo work.
Profiles keep each Claude Code or Codex session lean by materializing only the skills and MCPs needed for the current job.
soul list
soul use medusa-dev
cd profiles/medusa-dev/workspace && claudeStart with the docs hub at docs/profiles/ for the schema, inheritance model, scan-to-profile flow, and troubleshooting.
If you are an AI coding agent helping a human set this up, read AGENTS.md first — it explains the bootstrap contract (phase-by-phase, ask-before-network, verify-each-step) and points to the per-OS prompt your user should paste.
After bootstrap (~550 MB cold RAM per session):
| Layer | What it does |
|---|---|
| claude-mem plugin | Captures session observations passively; mem-search "topic" to recall across sessions |
| caveman plugin | /caveman for terse replies, /caveman-commit for Conventional Commit messages |
| RTK (CLI hook) | Filters shell-command outputs — 60-90% token savings on ls / git / cat |
| gbrain MCP | Personal knowledge wiki with embeddings, backlinks, timeline |
| excel-mcp-server | Native .xlsx read/write/format |
| office-word-mcp-server | Native .docx read/write |
claude-mem (passive) and gbrain (manual wiki) are complementary — both recommended.
soul/
├── skills/ 110+ Claude Code / Codex skills
│ ├── medusa/ building-with-medusa, storefront-best-practices, …
│ ├── codex-fleet/ bringup, dispatch, supervisors, panes, troubleshoot
│ ├── higgsfield/ generate, marketplace-cards, soul-id
│ ├── caveman/ caveman, caveman-commit, caveman-compress
│ └── ...
├── mcps/ MCP server snapshots + configs
│ ├── configs/ claude.sanitized.json, codex.sanitized.json, …
│ ├── mcps/ individual MCP server entries
│ └── plugins/ Claude Code plugin snapshots
└── setup/ paste-into-Claude-Code prompts (macos.md, linux.md, windows.md)
Each skill is a folder with SKILL.md (frontmatter + body) plus reference files. The frontmatter description is what the LLM matches against — write it as "when user says X, do Y".
To add a new skill: copy an existing one as a template, edit SKILL.md, drop it under skills/skills/<category>/<slug>/. The catalog regenerates on the next sync.