claude-code-setup

One-command installer for my Claude Code configuration — hooks, slash commands, skills, statusline, and global rules. Install once, get the exact setup I use every day. Optimised for low token usage.

Why this setup?

Stock Claude Code is powerful but neutral. This config turns it into a careful, cost-aware engineer with guardrails — the difference between an assistant that sounds confident and one that proves its work.

• 🛡️ Trustworthy by default. Built-in rules forbid the things that quietly burn you: never fakes or weakens tests, never suppresses errors with @ts-ignore / eslint-disable / swallowed catch, never claims "done" without verifying. Failures get surfaced, not hidden.
• 🔒 Catches mistakes before they ship. scan-secrets blocks any commit containing API keys or a real .env. stop-verify runs lint + typecheck on the files you changed at the end of every turn and won't let "done" slide if they fail. check-dep forces a real look at any new dependency (size, maintenance, lighter alternatives) before it's added.
• 🎯 Stays in scope. Only changes what you asked for — no drive-by refactors, no unrequested "improvements", no comments bolted onto code it didn't touch.
• 🧠 Remembers across sessions. The transcript-search MCP indexes your own past conversations locally, so Claude recalls prior decisions ("like we did last time") instead of saying "I don't remember" — and instead of you re-explaining context.
• 💸 Cheap on tokens. Caveman mode keeps full technical accuracy while cutting chatter ~75%; video defaults to audio-only transcription; verification hooks touch only changed files, not the whole repo. See Optimised for low token usage.
• 🎬 Understands video. The claude-video-vision MCP lets Claude watch and reason about video, not just text.
• 📊 Knows where you stand. A rich statusline shows git branch + dirty state, model and effort level, context-window %, and your 5-hour / 7-day rate-limit usage at a glance.
• ⚡ One-command, exact mirror. Installs everything, auto-installs its dependencies, and registers the MCP servers. It reproduces my config 1:1 — existing files are overwritten so you get an identical setup, with any file it replaces backed up to <file>.bak first. It never deletes files it doesn't manage. Idempotent — re-run any time to pull the latest.

Install

npx @antonior/claude-code-setup

That's it. The installer is idempotent — safe to re-run to pull updates.

Platforms

Built and tested on macOS. Linux and native Windows aren't officially supported yet — the hooks are bash and the installer uses Homebrew, so a full cross-platform pass is needed before I'd claim them.

Dependencies (auto-installed via Homebrew)

The installer installs these for you if missing:

| Tool | For | Notes | |------|-----|-------| | jq | hooks (JSON parsing) | Required — install aborts without it. | | python3 | transcript-search MCP + video-FPS hook | Best-effort — feature stays dormant if it can't install. | | ffmpeg | claude-video-vision MCP | Best-effort. | | tmux | optional cr tmux launch workflow (lives in shell config, not shipped) | Best-effort. | | git, eslint, tsc | commit-scan / lint / typecheck hooks | Optional — hooks no-op cleanly when absent. |

What it installs (into ~/.claude/)

• CLAUDE.md — global rules (trust/integrity, scope discipline, verification). Overwritten to match; any existing one is saved to CLAUDE.md.bak.
• settings.json + settings.local.json — permissions, hooks, effort/theme. Overwritten to match; any existing ones are saved to .bak.
• hooks/
  • caveman-activate.sh — terse response style on session start
  • scan-secrets.sh — blocks git commit if staged diff contains secrets/.env
  • check-dep.sh — prompts to research a dependency before adding it
  • eslint-fix.sh — auto eslint --fix on edited JS/TS files
  • stop-verify.sh — lint + typecheck changed files when a turn ends
  • notify-sound.sh — sound on notification/stop
• commands/ — /check-dep, /debug, /scan-secrets, /updateClaudeNpm slash commands
• skills/ — /mute, /unmute, and /caveman (with intensity levels)
• agents/ — custom subagents: Explore (fast read-only code search), Plan (architect/implementation plans), statusline-setup (statusline config)
• statusline.sh — git, model, context %, rate-limit statusline
• transcript-search/rag_lite.py — the transcript-search MCP engine
• MCP servers (registered via claude mcp add-json, user scope) — see below

Bundled MCP servers

Both are installed and auto-registered for you:

• transcript-search — full-text search over your own past Claude Code transcripts ("we talked about…", "like last time"). Pure-Python, stdlib only. The index is built locally on your machine from ~/.claude/projects/ on first run — nothing about your conversations is ever shipped in this package.
• claude-video-vision — lets Claude watch/analyse videos (frames via ffmpeg). Published as claude-video-vision on npm; run via npx.
  • Tip: when it asks how many FPS, answer 0 to use audio transcription only (no frame extraction) — much cheaper on tokens. Bump the FPS only when you actually need on-screen/visual detail.

Optimised for low token usage

This setup is deliberately token-lean:

• /caveman skill + caveman hook — ultra-compressed replies that keep full technical accuracy while cutting chatter (~75% fewer tokens), with lite/full/ultra intensity levels.
• Video → 0 FPS — analyse video by audio transcription only unless you explicitly need visual frames, avoiding expensive frame tokens.
• transcript-search over re-explaining — recalls past decisions from your own history instead of re-deriving context.
• Scoped, high-signal hooks — stop-verify and eslint-fix only act on files you actually changed, not whole-repo sweeps.

On a new machine — two one-time steps

The installer reproduces 100% of the configuration and behaviour. Two things are identity and secrets, not config, and can never live in a public package — you set them once:

1. Log into Claude Code (your account — you'd do this on any new machine anyway).
2. Give claude-video-vision your own API key if you use video analysis.

Shell workflow (optional, not shipped)

The package mirrors ~/.claude only, so anything in your shell config (~/.zshrc) is not shipped. Add it per machine. These two bits drive how I launch Claude:

Start Claude in bypass mode (skips per-tool permission prompts). Note: skipDangerousModePermissionPrompt in settings.json only suppresses the warning; this alias is what actually starts a session in bypass mode:

alias claude="claude --dangerously-skip-permissions"

cr — launch Claude inside a tmux session (named after the current directory, or pass a name: cr myname):

cr() {
  local session="${1:-$(basename "$PWD")}"
  session="${session//[.:]/_}"   # tmux treats . and : as target separators
  if tmux has-session -t "$session" 2>/dev/null; then
    tmux attach-session -t "$session"
  else
    tmux new-session -s "$session" -d -c "$PWD"
    tmux send-keys -t "$session" "claude --dangerously-skip-permissions --remote-control '$session'" Enter
    tmux attach-session -t "$session"
  fi
}

cr needs tmux, which the installer now installs best effort (see the dependencies table). After editing ~/.zshrc, run source ~/.zshrc or open a new terminal.

Maintaining (for me)

The package is an exact mirror of my own ~/.claude. To refresh it from my live config and publish:

• scripts/sync-from-local.sh rebuilds files/ from an allowlist of my config (CLAUDE.md, settings*.json, statusline.sh, hooks/, commands/, agents/, skills/, the caveman skill, the transcript-search engine). It never copies private or runtime data (history, transcripts, memory, the search index, sessions, caches, marketplace clones).
• The /updateClaudeNpm slash command runs that sync, scans for secrets, bumps the version, pushes to main, and publishes to npm — so a fresh npx @antonior/claude-code-setup on another machine reproduces my setup 1:1.

License

MIT © Antonio Radosav