Mr Future

Vol. I · No. 10October 2026

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

Feature

A quick start guide to coding with Pi and Qwen locally

Based on feedback from my Pi on a local model workshop: a one-slide quick start guide to running Pi and Qwen for local coding on a Mac.

By Bill McIntyre

Based on feedback from my workshop on Pi on a local model, I’ve created this one-page quick start guide to getting ramped up on running Pi and Qwen for local coding on a MacBook or macOS environment. The slide has all the links and pieces you need to get going very quickly.

Slide 2 of the Pi on a local model deck, “Cheat sheet: zero to a local agent”. Five steps: install the MTPLX app and remove its passwordless sudo rule; download Qwen3.8-27B (20 GB) and serve it on port 8000; install pi 1.3.0 with npm (Node 22.19 or newer); add pi-sandbox and pi-mtplx, pinned; register the model in pi and start working in your repo. Side boxes list what to put in AGENTS.md, pi settings (thinking set to low for routine edits, a 131,072-token context window and the built-in MCP extension off) and three skills from obra/superpowers.

The quick steps are here

  1. Install the MTPLX app from mtplx.com/download, then open a new terminal so the mtplx command is on your path. Its fan control adds a passwordless sudo rule. Remove it:

    sudo rm /etc/sudoers.d/mtplx-thermalforge

  2. Download the model (20 GB) and serve it:

    mtplx pull Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed
mtplx quickstart --port 8000

  3. Install pi (Node 22.19 or newer) from github.com/earendil-works/pi:

    npm install -g --ignore-scripts @earendil-works/pi-coding-agent@1.3.0

  4. Add the sandbox and the MTPLX bridge, pinned (pi-sandbox and pi-mtplx):

    pi install npm:pi-sandbox@0.7.2
pi install npm:pi-mtplx@0.2.0

  5. Register the model, restart, then work. Start pi, then /mtplx → Models, then /quit. Then, in your repo:

    cd my-repo && pi

    and pick the model with /model → mtplx.

First session recommendations

In AGENTS.md:

  • Run the tests and linter before a step is done
  • Keep PLAN.md current; commit each step
  • Blocked by the sandbox? Stop and say so

In pi’s settings:

  • /thinking low for routine edits
  • "contextWindow": 131072 in models.json
  • "extensions": ["-builtin:mcp"] in settings.json

Skills, copied into ~/.agents/skills from obra/superpowers:

  • test-driven-development
  • verification-before-completion
  • systematic-debugging

Tested on pi 1.3.0 and MTPLX 2.12 in October 2026.

The rest of the deck has some great tips on optimizing your context window, better ways of coding with this Qwen model, and a lot of other great info. But the new quick start guide will hopefully get you up and running.