The Pi-ed Piper

Just like many of you, I’ve been compelled to use “AI” ¹ for various tasks at work. It’s the currently-inescapable hype train ². At least it’s kind of helpful and interesting once corralled. After many months using different AI tools, I’ve settled on the Mimo model family and the Pi agent harness. I’m specifically using the pi-bolt fork of Pi. Let me show you how I’m using them currently.

Models

I only really use one subscription model, Mimo , as mentioned before. Currently it’s on v2.6 and has two variants, “flash” and “pro”. By default, every request and conversation uses the lighter flash variant. For things that benefit from heavier reasoning, like Open Design , I use the pro version. If you’re interested in trying Mimo, using this referral link will help us both out.

I’ve had good luck in the past using Ornith locally, if you’d like to use a local model. I haven’t used Ornith with Pi specifically ³ but I don’t know of any reason they shouldn’t play nice together.

What’s installed and why

I try to make sure each customization I add to any tool is useful. I don’t always succeed in goal but I think I have, so far, with Pi.

• bigpowers for brainstorming, planning, and general project organization. This is combined with pi-beads since I use beads for project management.
• pi-hermes-memory , billion-context-pi-lean , and pi-declutter to let me have long-running sessions without them becoming unreadable.
• tersio gives me rtk , caveman , and ponytail all at once. Each of these helps me to spend fewer tokens. I also combine them with pi-greybeard and soft-skills to make conversation even more concise.
• pi-jj and pi-jujutsu since I use jujutsu for version control. Models already know how to use git.
• pi-lens for code linting, language servers, etc.
• pi-ux , keystone , and pi-impeccable as design and quality gates.

There are some other packages installed but they’re for things like themes and status lines. If you’re interested in them, let me know and I’ll happily share.

Daily use

I always have my token-savers on; caveman, rtk, greybeard, and ponytail. The last two are interesting in that they reduce cost by reducing the amount of code that’s written or the number of allowed failures. Together they generally result in the least amount of code that will work. Between terse conversations and minimal code output, we’re already off to a good start.

For anything bigger than a couple of small tasks, I use a planning workflow that follows a few consistent steps — the same method I used before any of this and one that works just as well with pen and paper. The workflow itself, and how I hand it off to agents, lives in How I Plan Work With Agents .

Verdict?

This has, so far, been the most comfortable harness set up I’ve had yet. In the past, I’ve used Claude Code, OpenCode, and jcode . They’ve all been fine and fairly customizable but they’ve all felt, somewhat, like working on someone else’s computer. Pi, starting so barebones, has been more comfortable. It has that vim smell. I’m continuing to tweak my set up, of course, but this one seems like a winner.