Ask HN: How Are Developers Setting Up AI Coding Tools?

Summary

A Hacker News discussion compares how developers are incorporating AI into software development. One participant uses Claude Code, Codex, and OpenCode with Z.ai GLM models, mainly from iTerm2 on macOS or a terminal on Arch Linux; Emacs is occasionally used as the coding-agent environment, while JetBrains IDEs and VS Code are used more for reading and debugging. The setup is supplemented by custom tools that help agents plan multi-step work, stay on track, inspect generated code and documentation, restrict write access outside an assigned project, and collaborate through shared wikis, mailboxes, and GitHub CLI issues. The participant has also built regression-testing frameworks for agents. Web and desktop versions of Claude, ChatGPT, and Gemini are used for research, architecture, requirements, planning, design, and review. The developer is moving toward OpenCode with local models served by llama.cpp, Freetoken, or vLLM on GPUs with 12, 16, and 24 GB of VRAM, but reports that cloud LLMs remain faster and better so far. Recent Qwen 3.x mixture-of-experts models are described as promising. A second commenter reports using Claude Code integrated with VS Code and Visual Studio 2026 for development testing.