How to set up a local AI agent for coding and avoid dependency on rising cloud model prices

In response to rising prices and stricter rate limits for cloud-based AI coding models like Claude Code and GitHub Copilot, The Register explores using local models as a cost-saving alternative. The article focuses on Alibaba's Qwen3.6-27B model, which claims flagship coding power and can run on a 32 GB M-series Mac or 24 GB GPU. It provides a hands-on guide for deploying the model using Llama.cpp, specifying hyperparameters such as temperature=0.6, top_p=0.95, and a context window up to 262,144 tokens (limited by memory). The article then reviews three agent frameworks: Claude Code (works with local models by setting environment variables), Pi Coding Agent (lightweight open-source alternative with minimal system prompt), and Cline (VS Code extension with planning and action modes). Testing shows Qwen3.6-27B can one-shot a solar system web app and patch bugs in existing code. The article also covers security considerations, noting that Claude Code and Cline require human approval by default, while Pi Coding Agent operates with full autonomy and should be sandboxed. Overall, local models are not replacements for frontier models but are surprisingly capable.