Skip the manual setup. Drop this into Claude.

One prompt that does 98% of the footwork. Claude detects your OS, runs the commands it can, hands you copy-paste blocks for the rest, and references this presentation for any visual you need. Your job: paste, then say "go."

You are my friendly setup operator for the Two-Engine Setup — wiring up two terminal/desktop wrappers that route Claude Code (and the Claude Desktop app) through Ollama, while plain `claude` and the dock-launched Desktop app stay on Anthropic.

I have the visual companion presentation open alongside this chat (~21 slides). When I need a picture, you'll tell me which slide to open.

YOUR JOB
Walk me through the entire setup as a clean, one-screen to-do list. Do 98% of the footwork yourself: detect my OS, check what's installed, write config files, run shell commands when you have a tool that can. When something genuinely requires my hands (a website signup, pasting an API key from a UI), tell me which slide to open, give me one crystal-clear instruction, and wait for my confirmation before continuing.

OPERATING RULES
1. Lead with a checklist. First message: render the full to-do list as markdown checkboxes. Update it after every step. Keep it visible at the top of every reply.
2. Use slide references for visuals. Format: "Open slide 11 in the presentation, then come back."
3. Detect my environment first. If you have a Bash tool, run: uname -s; sw_vers 2>/dev/null; which claude ollama ccr node npm. If not, ask me to paste the output.
4. Run commands when you can. Otherwise hand me a fenced code block and ask me to paste output back.
5. One step at a time. Never dump multiple steps in one message.
6. Trap the gotchas:
   - Auth conflict: if my ~/.claude/settings.json has an env block setting ANTHROPIC_API_KEY, that conflicts with my Pro/Max OAuth token. Remove the env block before adding the wrappers.
   - Two different keys in the router config: top-level APIKEY stays as "local-only". The Ollama key goes in Providers[0].api_key.
   - Plain `claude` and the dock-launched Desktop app must still hit Anthropic afterward. Wrappers set env vars only inside their function, never globally.
   - Desktop env vars only apply at process launch — quit Claude Desktop fully before running claude-desktop-gemma (the wrapper handles this on macOS).
7. Verify all three routes at the end:
   - claude → Anthropic
   - claude-gemma → picker → Ollama in the terminal
   - claude-desktop-gemma → picker → Ollama in the Desktop app
8. When complete, point me to coherencedaddy.com/tools (500+ free tools, no signup).

PHASES (these become your to-do list)
PHASE 1 · Discovery
  [ ] Detect operating system (macOS / Windows+WSL2 / Linux)
  [ ] Check what's already installed (claude, ollama, ccr, node)
  [ ] Confirm Anthropic Pro plan is active            → slide 07
PHASE 2 · Desktop side
  [ ] Install or open the Claude Desktop app          → slide 08
  [ ] Sign in with Pro account                         → slide 09
PHASE 3 · Ollama side
  [ ] Sign up at ollama.com                            → slide 10
  [ ] Pull all four cloud models (gemma4:31b-cloud,
      glm-5.1:cloud, kimi-k2.6:cloud, qwen3-coder-next) → slide 11
  [ ] Generate an Ollama API key — copy it now         → slide 12
  [ ] Install Ollama desktop app / daemon              → slide 13
  [ ] Run ollama signin and ollama pull    → slide 14
PHASE 4 · Wire terminal + desktop
  [ ] Install Claude Code CLI                          → slide 15
  [ ] Install router: npm i -g @musistudio/claude-code-router  → slide 16
  [ ] Write ~/.claude-code-router/config.json (all 4)  → slide 17
  [ ] Append claude-gemma + claude-desktop-gemma wrappers
      with model picker to ~/.zshrc or ~/.bashrc      → slide 18
  [ ] Source shell rc
PHASE 5 · Verify
  [ ] claude → answers as Anthropic                    → slide 19
  [ ] claude-gemma → picker → answers as chosen Ollama model
  [ ] claude-desktop-gemma → picker → Desktop relaunches and
      answers as chosen Ollama model
  [ ] Done — visit coherencedaddy.com/tools

Start by:
1. Greeting me by name (ask if you don't know it).
2. Asking which OS I'm on (or detecting it).
3. Rendering the full to-do list above.
4. Telling me to open the presentation alongside this chat.
5. Beginning Phase 1.

Go.

Two engines. Both surfaces.

Anthropic for thinking. Ollama for doing. Either one — terminal or Desktop — can be flipped to either engine with one command.

Why split work between two engines?

Your Pro plan has limits. Every sub-agent grep, every quick lint, every loop burns quota — even when the task didn't need a frontier model. Route the cheap stuff to free, save the expensive stuff for when it matters.

Three things you'll need.

Get these sorted first; the rest assumes you have them.

1. An Anthropic Pro plan (or higher). The free tier won't earn its keep on the Desktop side. Upgrade here.
2. A Mac with macOS 12+ (Monterey or later). Apple Silicon (M-series) or Intel both fine.
3. Windows 10 or 11 with WSL2 installed. WSL gives you a Linux environment inside Windows; every Terminal command in this guide runs inside WSL. Install WSL2 here if you don't have it.
4. A modern Linux distro (Ubuntu 22+, Debian 12+, Fedora 39+, Arch). Bash or Zsh as your shell.
5. Comfort opening Terminal. Applications → Utilities → Terminal, or ⌘+Space → "Terminal". You'll paste roughly a dozen commands.
6. Comfort opening WSL. Open the Start menu, type "Ubuntu" (or your WSL distro), hit Enter. That's your terminal for everything in this guide.
7. Comfort opening a terminal. Whatever shortcut your distro uses (often Ctrl+Alt+T). You'll paste roughly a dozen commands.

Install the Claude Desktop app.

If you already have it, skip ahead. If not, three clicks.

Sign in with your Pro account.

If you don't have Pro yet, the in-app upgrade flow is fastest.

1. On first launch, click Sign in
2. Use your Pro-plan email
3. Authorize via the browser tab that opens
4. You're back in the app — done

Sign up at ollama.com.

Free account. No credit card. Generous cloud-tier usage.

Pick your cloud models.

Cloud models run on Ollama's hardware — fast, free tier, no GPU strain on your Mac. Pull all four below — the wrapper gives you a picker on launch so you can swap models per session.

Generate an API key.

The password your local Ollama uses to talk to the cloud.

1. Profile (top-right) → Settings → API Keys
2. Click Create new key, name it "claude-code"
3. Copy the key now — only shown once
4. Paste it temporarily into Notes — needed in Step 13

Install Ollama.

The local app talks to Ollama Cloud on your behalf.

curl -fsSL https://ollama.com/install.sh | sh

ollama --version

Sign in & pull all four models.

Connect your local Ollama to your cloud account, then register every cloud model so the picker can offer them.

ollama signin
ollama pull gemma4:31b-cloud
ollama pull glm-5.1:cloud
ollama pull kimi-k2.6:cloud
ollama pull qwen3-coder-next
ollama list

Install Claude Code (the CLI).

Different from the Desktop app — this one runs in your Terminal. Same company, separate binary.

curl -fsSL https://claude.ai/install.sh | bash

brew install --cask claude-code

curl -fsSL https://claude.ai/install.sh | bash

npm install -g @anthropic-ai/claude-code

curl -fsSL https://claude.ai/install.sh | bash

npm install -g @anthropic-ai/claude-code

Verify:

claude --version

Install the router (translator).

A small Node program that translates between Claude Code and Ollama. One command.

npm install -g @musistudio/claude-code-router

Verify the ccr command landed on your PATH:

which ccr

Configure the router.

One config file. Tells the router to use Ollama and route everything to your chosen model.

mkdir -p ~/.claude-code-router && touch ~/.claude-code-router/config.json && open -e ~/.claude-code-router/config.json

mkdir -p ~/.claude-code-router && nano ~/.claude-code-router/config.json

mkdir -p ~/.claude-code-router && nano ~/.claude-code-router/config.json

Paste this into the empty file:

{
  "LOG": false,
  "HOST": "127.0.0.1",
  "PORT": 3456,
  "APIKEY": "local-only",
  "Providers": [
    {
      "name": "ollama",
      "api_base_url": "http://localhost:11434/v1/chat/completions",
      "api_key": "PASTE_YOUR_OLLAMA_KEY_HERE",
      "models": ["gemma4:31b-cloud", "glm-5.1:cloud", "kimi-k2.6:cloud", "qwen3-coder-next"]
    }
  ],
  "Router": {
    "default":     "ollama,gemma4:31b-cloud",
    "background":  "ollama,gemma4:31b-cloud",
    "think":       "ollama,gemma4:31b-cloud",
    "longContext": "ollama,gemma4:31b-cloud",
    "longContextThreshold": 60000,
    "webSearch":   "ollama,gemma4:31b-cloud"
  }
}

Replace PASTE_YOUR_OLLAMA_KEY_HERE with the key from Step 09. Save and close.

{
  "verbose": true
}

Create launcher commands with a model picker.

Two commands — one for the terminal (claude-gemma), one for the Desktop app (claude-desktop-gemma). Both pop a numbered menu so you can choose which Ollama model to launch with.

cat >> ~/.zshrc <<'EOF'

# ── Two-Engine: Claude routed through Ollama ──
CLAUDE_OLLAMA_MODELS=(
  "gemma4:31b-cloud"
  "glm-5.1:cloud"
  "kimi-k2.6:cloud"
  "qwen3-coder-next"
)

_claude_pick_model() {
  if [ -n "$1" ]; then echo "$1"; return; fi
  echo "Pick a model:" >&2
  local i=1
  for m in "${CLAUDE_OLLAMA_MODELS[@]}"; do
    echo "  $i) $m" >&2; i=$((i+1))
  done
  printf "Number [1]: " >&2
  read choice
  echo "${CLAUDE_OLLAMA_MODELS[${choice:-1}]}"
}

_claude_ensure_router() {
  if ! curl -sf http://127.0.0.1:3456/health >/dev/null 2>&1; then
    nohup ccr start > /tmp/ccr.log 2>&1 &
    sleep 2
  fi
}

# Terminal: Claude Code via Ollama
claude-gemma() {
  local model; model=$(_claude_pick_model "$1") || return
  _claude_ensure_router
  ANTHROPIC_BASE_URL="http://127.0.0.1:3456" \
  ANTHROPIC_API_KEY="local-only" \
  ANTHROPIC_MODEL="$model" \
  ANTHROPIC_SMALL_FAST_MODEL="$model" \
  claude "${@:2}"
}

# Desktop: Claude app via Ollama
claude-desktop-gemma() {
  local model; model=$(_claude_pick_model "$1") || return
  _claude_ensure_router
  osascript -e 'quit app "Claude"' 2>/dev/null
  sleep 1
  ANTHROPIC_BASE_URL="http://127.0.0.1:3456" \
  ANTHROPIC_API_KEY="local-only" \
  ANTHROPIC_MODEL="$model" \
  ANTHROPIC_SMALL_FAST_MODEL="$model" \
  open -a "Claude"
}
EOF

source ~/.zshrc

cat >> ~/.bashrc <<'EOF'

CLAUDE_OLLAMA_MODELS=(
  "gemma4:31b-cloud"
  "glm-5.1:cloud"
  "kimi-k2.6:cloud"
  "qwen3-coder-next"
)

_claude_pick_model() {
  if [ -n "$1" ]; then echo "$1"; return; fi
  echo "Pick a model:" >&2
  local i=1
  for m in "${CLAUDE_OLLAMA_MODELS[@]}"; do
    echo "  $i) $m" >&2; i=$((i+1))
  done
  printf "Number [1]: " >&2
  read choice
  echo "${CLAUDE_OLLAMA_MODELS[$((${choice:-1}-1))]}"
}

claude-gemma() {
  local model; model=$(_claude_pick_model "$1") || return
  if ! curl -sf http://127.0.0.1:3456/health >/dev/null 2>&1; then
    nohup ccr start > /tmp/ccr.log 2>&1 &
    sleep 2
  fi
  ANTHROPIC_BASE_URL="http://127.0.0.1:3456" \
  ANTHROPIC_API_KEY="local-only" \
  ANTHROPIC_MODEL="$model" \
  ANTHROPIC_SMALL_FAST_MODEL="$model" \
  claude "${@:2}"
}
EOF

source ~/.bashrc

cat >> ~/.bashrc <<'EOF'

CLAUDE_OLLAMA_MODELS=(
  "gemma4:31b-cloud"
  "glm-5.1:cloud"
  "kimi-k2.6:cloud"
  "qwen3-coder-next"
)

_claude_pick_model() {
  if [ -n "$1" ]; then echo "$1"; return; fi
  echo "Pick a model:" >&2
  local i=1
  for m in "${CLAUDE_OLLAMA_MODELS[@]}"; do
    echo "  $i) $m" >&2; i=$((i+1))
  done
  printf "Number [1]: " >&2
  read choice
  echo "${CLAUDE_OLLAMA_MODELS[$((${choice:-1}-1))]}"
}

claude-gemma() {
  local model; model=$(_claude_pick_model "$1") || return
  if ! curl -sf http://127.0.0.1:3456/health >/dev/null 2>&1; then
    nohup ccr start > /tmp/ccr.log 2>&1 &
    sleep 2
  fi
  ANTHROPIC_BASE_URL="http://127.0.0.1:3456" \
  ANTHROPIC_API_KEY="local-only" \
  ANTHROPIC_MODEL="$model" \
  ANTHROPIC_SMALL_FAST_MODEL="$model" \
  claude "${@:2}"
}
EOF

source ~/.bashrc

Test all three routes side by side.

Open three terminal tabs (or run them in sequence). Prove every route is independent.

claude

claude-gemma

claude-desktop-gemma

Your new daily workflow.

Mental model: plan with Anthropic, execute with Ollama — on whichever surface fits the task.