Use local when...

• Privacy matters — code never leaves your machine
• You're offline — planes, cafes, bad wifi
• Latency is critical — autocomplete, quick lookups
• Experimenting freely — no token bill anxiety
• Running many requests — batch processing, CI pipelines

Use API when...

• You need frontier quality — Opus/Sonnet for hard problems
• Long context — 200K tokens isn't possible locally
• Complex reasoning — multi-step planning, architecture
• Speed at scale — cloud GPUs are faster than your laptop

# Install
brew install ollama

# Pull and run a model
ollama run llama3.3
ollama run deepseek-r1:8b
ollama run qwen2.5-coder:32b

# List installed models
ollama list

# Serve as API (OpenAI-compatible)
ollama serve
# http://localhost:11434/v1/chat/completions

Performance on Apple Silicon

M3 Pro (18GB) comfortably runs up to ~14B. M3 Max (36-96GB) handles 32B-70B.

# ~/modelfiles/code-reviewer
FROM qwen2.5-coder:14b

PARAMETER temperature 0.3
PARAMETER num_ctx 32768

SYSTEM """
You are a senior code reviewer.
Focus on:
- Security vulnerabilities
- Performance issues
- Readability improvements
Be concise. No praise, just actionable
feedback with line references.
"""

# Create and run it
ollama create code-reviewer \
  -f ~/modelfiles/code-reviewer
ollama run code-reviewer

Key Parameters

• temperature — 0 = deterministic, 1+ = creative
• num_ctx — context window size in tokens
• top_p — nucleus sampling threshold
• seed — fix random seed for reproducible output

# Three levels of memory

~/.claude/CLAUDE.md          # Global
# Your preferences across ALL projects
# "I prefer TypeScript, use pnpm, etc."

./CLAUDE.md                  # Project
# Shared with team via git
# "Use our API conventions, run tests
#  with vitest, deploy to Cloudflare"

./src/api/CLAUDE.md          # Subdirectory
# Auto-loaded when Claude reads
# files in this directory

Key Slash Commands

Skills — reusable knowledge

# .claude/skills/my-skill/SKILL.md
---
name: api-builder
description: Build REST APIs using our
  Hono + D1 + Cloudflare Workers stack
---

# API Builder
When building API endpoints:
- Use Hono router with typed middleware
- Validate inputs with Zod schemas
- Use D1 for database, KV for cache
- Write tests with vitest
...

Hooks — automated actions

// .claude/settings.json
{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit",
      "command": "prettier --write $FILE"
    }]
  }
}

• PostToolUse — auto-format after every file edit
• PreToolUse — validate before destructive actions
• Runs shell commands with context about what Claude just did

Tailscale is a WireGuard-based mesh VPN. Every device gets a stable IP and DNS name. No port forwarding, no firewall rules, no VPN server to manage.

# Install
brew install tailscale
tailscale up

# SSH to your VPS by name
ssh my-vps

# Access Ollama on your home Mac
# from anywhere:
curl http://macbook:11434/v1/models

# Share a dev server publicly
tailscale funnel 3000

Key Features

• MagicDNS — ssh my-vps instead of remembering IPs
• Funnel — expose local ports to the internet with HTTPS
• ACLs — control who can access what
• Exit nodes — route traffic through any of your machines

Provider Picks

Hetzner wins for personal setups — EU-based, dirt cheap, great specs.

What to run there

# Persistent Claude Code via tmux
ssh my-vps
tmux new -s agent

# Inside tmux: Claude Code keeps running
# even when you disconnect
claude

# Detach: Ctrl-B, then D
# Reattach later:
tmux attach -t agent

• Cron-triggered agents (daily reports, monitoring)
• MCP servers accessible from Claude Code anywhere
• Webhook handlers for GitHub, Slack, etc.
• Long-running background tasks

• Install Ollama — brew install ollama
• Pull a model — ollama run llama3.3
• Create a Modelfile — custom code reviewer persona
• Install Tailscale — brew install tailscale && tailscale up
• Set up CLAUDE.md — claude /init in your project
• Create a skill — build a reusable skill for your workflow
• Add a hook — auto-format on file writes
• SSH to VPS — start a persistent tmux session