setup.sh

#!/bin/bash
# Claude Code Local — One-command setup
# Apple Silicon only. Installs MLX, downloads a model from the lineup,
# and creates a desktop launcher that runs Claude Code 100% on-device.
#
# Usage: bash setup.sh

set -e

echo ""
echo "╔══════════════════════════════════════════════════╗"
echo "║     Claude Code Local — Setup                    ║"
echo "║     Pick your fighter. Run AI on your Mac.       ║"
echo "╚══════════════════════════════════════════════════╝"
echo ""

# ── System detection ──────────────────────────────────────────
MEM_GB=$(sysctl -n hw.memsize 2>/dev/null | awk '{print int($1/1073741824)}')
CHIP=$(sysctl -n machdep.cpu.brand_string 2>/dev/null || echo 'Apple Silicon')

echo "Detected: $CHIP"
echo "Memory:   ${MEM_GB} GB"
echo ""

if [[ $(uname -m) != "arm64" ]]; then
  echo "ERROR: This requires Apple Silicon (M1 or later)."
  exit 1
fi

# ── Homebrew ──────────────────────────────────────────────────
if ! command -v brew &>/dev/null; then
  echo "Installing Homebrew..."
  /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
  eval "$(/opt/homebrew/bin/brew shellenv)"
fi

# ── Python 3.12 + MLX ─────────────────────────────────────────
if ! command -v python3.12 &>/dev/null; then
  echo "Installing Python 3.12 (required for MLX)..."
  brew install python@3.12
fi

MLX_VENV="$HOME/.local/mlx-server"
if [ ! -d "$MLX_VENV" ]; then
  echo "Creating MLX virtualenv at $MLX_VENV..."
  python3.12 -m venv "$MLX_VENV"
fi

echo "Installing mlx-lm into virtualenv..."
"$MLX_VENV/bin/pip" install --quiet --upgrade pip
"$MLX_VENV/bin/pip" install --quiet --upgrade mlx-lm

# ── Pick a model from the lineup ──────────────────────────────
echo ""
echo "Selecting a model from the lineup for your ${MEM_GB} GB Mac..."
if [ "$MEM_GB" -ge 96 ]; then
  # Passed a Claude Code smoke test on an M5 Max, 2026-09-19 (wrote hello.py, ran it, 38 s).
  MODEL_ID="lmstudio-community/Qwen3.8-27B-MLX-8bit"
  MODEL_LABEL="Qwen 3.8 27B, 8-bit (~29 GB)"
  MODEL_TIER="🟣 max"
elif [ "$MEM_GB" -ge 64 ]; then
  MODEL_ID="divinetribe/gemma-4-31b-it-abliterated-4bit-mlx"
  MODEL_LABEL="Gemma 4 31B Abliterated, 4-bit (~18 GB)"
  MODEL_TIER="🟢 fast"
elif [ "$MEM_GB" -ge 32 ]; then
  MODEL_ID="divinetribe/gemma-4-12B-it-abliterated-4bit-mlx-text"
  MODEL_LABEL="Gemma 4 12B Abliterated, 4-bit (~11 GB)"
  MODEL_TIER="🟢 fast"
elif [ "$MEM_GB" -ge 16 ]; then
  MODEL_ID="divinetribe/Hermes-4-14B-abliterated-4bit-mlx"
  MODEL_LABEL="Hermes 4 14B Abliterated, 4-bit (~8 GB)"
  MODEL_TIER="🟡 laptop"
else
  # Loads and chats, but in a Claude Code smoke test (2026-09-19) it claimed to run a
  # file it never wrote. Small Macs get it anyway; tool use will be unreliable.
  MODEL_ID="mlx-community/gemma-4-e4b-it-4bit"
  MODEL_LABEL="Gemma 4 E4B, 4-bit (~5 GB, may struggle with tool calls)"
  MODEL_TIER="🟠 small"
fi

echo "Selected: $MODEL_TIER  $MODEL_LABEL"
echo "Model ID: $MODEL_ID"
echo ""

# ── Download model ────────────────────────────────────────────
echo "Downloading $MODEL_ID (one time, 5-30 GB)..."
"$MLX_VENV/bin/python3" - <<PY
from mlx_lm.utils import load
load("$MODEL_ID")
print("Done.")
PY

# ── Install MLX server ────────────────────────────────────────
# We install it as a symlink into this repo, not a copy. That way if you edit
# proxy/server.py in the repo (to fix a bug, add a feature, pull a PR), the
# change takes effect on the running server after the next restart — with no
# risk of the running copy silently drifting from the version in git.
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
SERVER_DIR="$HOME/.local/mlx-native-server"
mkdir -p "$SERVER_DIR"
ln -sf "$SCRIPT_DIR/proxy/server.py" "$SERVER_DIR/server.py"
echo "MLX server installed (symlink) → $SERVER_DIR/server.py -> $SCRIPT_DIR/proxy/server.py"

# ── Create desktop launcher ───────────────────────────────────
CLAUDE_BIN=$(which claude 2>/dev/null || echo "$HOME/.local/bin/claude")
if [ ! -f "$CLAUDE_BIN" ]; then
  echo ""
  echo "WARNING: Claude Code not found. Install it with:"
  echo "  npm install -g @anthropic-ai/claude-code"
  echo ""
  CLAUDE_BIN="\$HOME/.local/bin/claude"
fi

LAUNCHER="$HOME/Desktop/Claude Local.command"
cat > "$LAUNCHER" <<LAUNCH
#!/bin/bash
# Claude Code — Local AI ($MODEL_LABEL)
CLAUDE_BIN="$CLAUDE_BIN"
MLX_PYTHON="$MLX_VENV/bin/python3"
MLX_SERVER="$SERVER_DIR/server.py"

if ! lsof -i :4000 >/dev/null 2>&1; then
  MLX_MODEL="$MODEL_ID" "\$MLX_PYTHON" "\$MLX_SERVER" >/tmp/mlx-server.log 2>&1 &
  echo "  Loading $MODEL_LABEL on MLX..."
  while ! curl -s http://localhost:4000/health 2>/dev/null | grep -q "ok"; do
    sleep 2
  done
fi

clear
echo ""
echo "  → Claude Code with LOCAL AI"
echo "  → $MODEL_LABEL"
echo "  → 100% on-device, no cloud, no API fees"
echo ""

ANTHROPIC_BASE_URL=http://localhost:4000 \\
ANTHROPIC_API_KEY=sk-local \\
exec "\$CLAUDE_BIN" --model claude-sonnet-4-6
LAUNCH

chmod +x "$LAUNCHER"

# ── "keep going": continue a Claude conversation on a free or local model ──
mkdir -p "$HOME/.local/bin"
ln -sf "$SCRIPT_DIR/bin/keepgoing" "$HOME/.local/bin/keepgoing"
ZSHRC="$HOME/.zshrc"
KEEP_LINE="source \"$SCRIPT_DIR/scripts/keepgoing-shell.sh\""
if ! grep -qF "$KEEP_LINE" "$ZSHRC" 2>/dev/null; then
  printf '\n# Claude Code Local: type "keep going" when Claude runs out of credits\n%s\n' "$KEEP_LINE" >> "$ZSHRC"
fi
echo "keep going installed → ~/.local/bin/keepgoing (open a new terminal to use it)"
# Optional: run keep going by itself when Claude hits a usage limit. setup.sh never edits
# ~/.claude/settings.json, so it only prints the snippet (docs/AUTO-HANDOFF.md explains it).
echo ""
echo "Want keep going to start by itself when you hit a Claude usage limit? Add this to"
echo "~/.claude/settings.json (merge it into an existing \"hooks\" block if you have one):"
cat <<HOOK
  "hooks": {
    "StopFailure": [
      { "matcher": "rate_limit",
        "hooks": [ { "type": "command", "command": "python3 $SCRIPT_DIR/scripts/limit-handoff.py" } ] }
    ]
  }
HOOK

# ── Optional: iMessage / Screen-to-Phone tools ────────────────
echo ""
echo "Checking for optional iMessage phone-control tools..."
CLAUDE_DIR="$HOME/.claude"
mkdir -p "$CLAUDE_DIR"

PHONE_SCRIPTS=(
  "scripts/imessage-send.sh"
  "scripts/imessage-send-image.sh"
  "scripts/imessage-send-video.sh"
  "scripts/imessage-toggle.sh"
  "scripts/imessage-receive.sh"
)

MISSING=0
for s in "${PHONE_SCRIPTS[@]}"; do
  if [ -f "$SCRIPT_DIR/$s" ]; then
    cp "$SCRIPT_DIR/$s" "$CLAUDE_DIR/$(basename $s)"
    chmod +x "$CLAUDE_DIR/$(basename $s)"
  else
    MISSING=1
  fi
done

if [ "$MISSING" -eq 0 ]; then
  echo "  ✅ Phone tools installed to ~/.claude/"
  if [ -f "$SCRIPT_DIR/config.example.sh" ] && [ ! -f "$SCRIPT_DIR/config.sh" ]; then
    cp "$SCRIPT_DIR/config.example.sh" "$SCRIPT_DIR/config.sh"
    echo "  → Edit config.sh with your iPhone number + Apple ID"
  fi
  cp "$SCRIPT_DIR/config.sh" "$CLAUDE_DIR/screen-to-phone-config.sh" 2>/dev/null || true
else
  echo "  ℹ️  Phone scripts not bundled — clone them separately if you want phone control:"
  echo "     https://github.com/nicedreamzapp/claude-screen-to-phone"
fi

echo ""
echo "╔══════════════════════════════════════════════════╗"
echo "║     Setup complete!                              ║"
echo "╠══════════════════════════════════════════════════╣"
echo "║                                                  ║"
echo "║  Model:    $MODEL_ID"
echo "║  Server:   $SERVER_DIR/server.py"
echo "║  Launcher: ~/Desktop/Claude Local.command"
echo "║                                                  ║"
echo "║  Double-click 'Claude Local' on your Desktop     ║"
echo "║  to start coding with local AI.                  ║"
echo "║                                                  ║"
echo "║  Want a different model? See launchers/ for      ║"
echo "║  Gemma 4 Code, Qwen 3.8 Code, Browser Agent,     ║"
echo "║  and Narrative Gemma launchers.                  ║"
echo "║                                                  ║"
echo "╚══════════════════════════════════════════════════╝"
echo ""