Are you paying per token for every line of code you write? There's a better way. AI-assisted development has changed how software gets built, but it's also created a new kind of tax. A single coding session can cost $5 to $20. A team of ten running agentic workflows burns through $10,000+ a month. And your proprietary code, your unreleased features, your client data? It's all being processed on someone else's infrastructure. The Local-First AI Developer shows you how to take back control of your costs, your privacy, and your workflow. The 80/20 principle that changes everything This isn't an anti-cloud manifesto. About 80% of daily AI work (code completion, test generation, refactoring, debugging) runs beautifully on models you own, on hardware you control, at near-zero cost. The remaining 20% genuinely benefits from frontier cloud models. Both tiers earn their place. This book shows you how to build a workflow that uses each one where it actually matters. What you'll learn Starting with how LLMs actually work (tokens, context windows, quantisation, sampling), you'll build a complete local AI development environment from the ground up: The real economics. Per-token costs, break-even calculations, and exactly when local hardware pays for itself versus when cloud still wins. Hardware that fits your budget. Benchmarks covering Apple Silicon (M1 through M4 Ultra), NVIDIA GPUs (RTX 3090 through 5090), and AMD. Why memory bandwidth determines inference speed, and why a Mac Studio often outperforms a gaming GPU costing twice as much. Your first model in 30 minutes. Choosing between Llama, Qwen, Gemma, DeepSeek, and Phi. A practical decision framework covering quantisation trade-offs, context requirements, and RAM budgets, with a cheat sheet matching hardware tiers to task types. Complete privacy. Code never leaves your machine. For freelancers, agencies, and anyone working under NDAs, local AI isn't just cheaper. It's the only responsible choice. 10 AI coding tools in depth. Dedicated chapters on Cursor, GitHub Copilot, Claude Code, Codex CLI, Aider, OpenCode, Continue.dev, Pi, RepoPrompt, and OpenClaw. Practical guidance on strengths, limitations, and when to reach for each one. Intelligent local-cloud routing. Commodity tasks run locally at zero cost while complex reasoning escalates to cloud models automatically. A real case study. RON, a multi-platform compiler generating native Swift, Kotlin, and C# from a single DSL. 11 platform targets, built by one