Vol. I · No. 10October 2026 · Boulder, Colorado

Mr Future

Bill McIntyre's personal blog on AI/ML engineering, learning, and experiments.

“The best way to predict the future is to invent it.” — Alan Kay, Xerox PARC, 1971

Broadcast

Pi on a local model: a field guide for Claude Code users

You can run a genuinely useful coding agent entirely on a laptop. A workshop on the Pi coding agent with Qwen3.8-27B, for developers who use Claude Code every day.

By Bill McIntyre

This was the first of two sessions at the October AI/ML Engineering meetup on local and sovereign AI. The question: if you use Claude Code with a frontier model every day, what does it take to get real work done with a coding agent running entirely on your own laptop?

The deck walks through it in order:

  • The model. Why a 27B model when giants are a click away, and what makes Qwen3.8-27B stand out.
  • The runtime. Two engines compared: LM Studio and MTPLX.
  • Pi as a coding environment. How it sits next to Claude Code, and what it’s like to drive.
  • Making it safe. Claude Code’s safety in four layers, and a host sandbox for Pi.
  • Prompting a smaller model. Tests as the steering wheel (red, then green), living within a smaller context window, and the prompt habits that suit a 27B.
  • Getting started. A cheat sheet from zero to a working local agent, plus my settings.

The slides work in any browser: arrow keys to move, G to see every slide, N for speaker notes. New to this? Start with slide 02, the cheat sheet.

The workshop’s companion piece is COLD STORAGE, a text adventure written entirely by the local model.