🤖 Claude Code — Complete Setup Guide

Run Claude Code with local Ollama models (free) or authenticate via Claude Pro subscription / Anthropic API.
Covers macOS M-Series, NVIDIA RTX 40-series, Blackwell, and DGX Spark hardware.

---

📋 What's Inside

File	Description
Claude_Code_Complete_Setup_Guide.pdf	Full illustrated guide with tables, hardware specs, and troubleshooting
scripts/setup_ollama.sh	One-shot setup script for macOS/Linux — Ollama + Claude Code
scripts/reset_claude_code.sh	Clean-slate reset — removes all Ollama/OpenClaw/dummy-key config

---

⚡ Quick Start (Choose Your Path)

Option 1 — Free Local Models via Ollama

# 1. Install Ollama
brew install ollama          # macOS
curl -fsSL https://ollama.com/install.sh | sh   # Linux

# 2. Pull a model (see hardware guide below)
ollama pull qwen3-coder

# 3. Install Claude Code
curl -fsSL https://claude.ai/install.sh | bash

# 4. Set env vars in ~/.zshrc
echo 'export ANTHROPIC_BASE_URL="http://localhost:11434"' >> ~/.zshrc
echo 'export ANTHROPIC_AUTH_TOKEN="ollama"' >> ~/.zshrc
echo 'export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-4.7-flash"' >> ~/.zshrc
echo 'export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3-coder"' >> ~/.zshrc
echo 'export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen3-coder"' >> ~/.zshrc
source ~/.zshrc

# 5. Launch
ollama serve &
claude

Option 2 — Claude Pro / Max Subscription (Real Claude Models)

# No Ollama needed. Just install Claude Code and login.
curl -fsSL https://claude.ai/install.sh | bash
claude login    # Opens browser — sign in with your claude.ai account

Option 3 — Anthropic API Key

curl -fsSL https://claude.ai/install.sh | bash
echo 'export ANTHROPIC_API_KEY="sk-ant-api03-your-key-here"' >> ~/.zshrc
source ~/.zshrc
claude

---

🖥️ Model Recommendations by Hardware

Apple Silicon Macs (M1 / M2 / M3 / M4)

RAM	Chip	Recommended Models
8 GB	M1/M2/M3/M4 base	glm-4.7-flash, qwen2.5-coder:7b
16 GB	M1/M2/M3/M4 Pro	qwen3-coder, deepseek-coder:13b
24 GB	M1/M2/M3 Max · M4 Max	qwen3-coder, gpt-oss:20b, deepseek-r1:32b
48 GB+	M1/M2/M3 Ultra · M4 Ultra	deepseek-r1:70b, qwen3-coder:72b

NVIDIA GPUs (RTX 40-Series & Blackwell)

GPU	VRAM	Recommended Models
RTX 4060	8 GB	glm-4.7-flash, qwen2.5-coder:7b
RTX 4060 Ti	16 GB	qwen3-coder, deepseek-coder:13b
RTX 4070 / Ti / Super	12–16 GB	qwen3-coder, gpt-oss:20b
RTX 4080 / Super	16 GB	gpt-oss:20b, deepseek-r1:32b (Q4)
RTX 4090 Workstation	24 GB	deepseek-r1:32b, llama3.3:70b (Q4)
RTX 6000 Ada / Blackwell	48–96 GB	deepseek-r1:70b, qwen3-coder:72b
Multi-GPU RTX 6000 (NVLink)	96–192 GB	Full 70B+ in FP16
NVIDIA DGX Spark (GB10)	128 GB	All 70B+ models, multiple concurrent agents

Model Tier Mapping (Required for Claude Code)

Claude Code uses 3 internal tiers. You must map all three:

# Example for 16 GB hardware (Mac 16GB / RTX 4060 Ti)
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-4.7-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3-coder"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen3-coder"

Hardware	Haiku	Sonnet	Opus
8 GB	glm-4.7-flash	qwen2.5-coder:7b	qwen2.5-coder:7b
16 GB	glm-4.7-flash	qwen3-coder	qwen3-coder
24 GB	qwen2.5-coder:7b	qwen3-coder	deepseek-r1:32b
48 GB+	qwen2.5-coder:7b	qwen3-coder:72b	deepseek-r1:70b

---

♻️ Reset to Clean Slate

If you had a previous Ollama / OpenClaw / dummy-key setup:

chmod +x scripts/reset_claude_code.sh
./scripts/reset_claude_code.sh

This removes all Anthropic env vars, logs out, and resets Claude Code config.
After reset, run claude login to authenticate with Claude Pro.

---

⚠️ Common Gotchas

• ANTHROPIC_API_KEY set + Claude Pro subscription → Claude Code uses the API key and bills per-token, ignoring your subscription. Remove the env var and use claude login instead.
• Not setting all 3 model tiers → causes model not found errors mid-session.
• Ollama not running → ConnectionRefused. Always run ollama serve before launching Claude.
• VS Code not picking up changes → must fully quit (Cmd+Q) and reopen, not just close the window.

---

📚 Resources

• Claude Code Docs
• Ollama
• Anthropic Console
• Ollama Anthropic Compatibility Docs

---

🙌 Contributing

Found a better model for a specific hardware config? Open a PR or raise an issue — hardware landscape changes fast and community updates are welcome.

---

📄 License

MIT — free to use, share, and adapt with attribution.