A comprehensive tutorial on setting up a fully local coding agent stack using open-weight LLMs as an alternative to proprietary services like Claude Code and Codex. Covers installing Ollama to serve models locally (primarily Qwen3.6 35B-A3B), connecting it to three harnesses (Qwen-Code, Codex CLI, and Claude Code), performing speed and benchmark assessments, auditing agent codebases for security and telemetry risks, and configuring privacy settings. Includes token usage comparisons across harnesses (Claude Code uses far more input tokens), a cross-machine SSH tunnel setup for running the model on a separate DGX Spark while coding on a Mac, and benchmark results showing Qwen3.6 and North Mini Code outperform smaller models like Gemma 4 E2B for agentic tool-use tasks.
