i move pixels Chris Devine

Experiments

Robots that tell me what Claude Code is doing

A desktop widget for keeping track of several Claude Code sessions at once.

The problem

I run several projects through Claude Code at once, and I kept losing track. I would forget what a session was working toward. I would miss a terminal window that had been waiting on my approval, or one that had finished and could have started its next task an hour ago.

I tried the status tools that exist and came away underwhelmed. They told me a session was busy. I wanted to know what it was doing, and what it needed from me.

What it does

A small floating window on my Mac. Every Claude Code session gets its own robot, with the project’s name and a line underneath saying what it is doing right now.

Two sessions at a glance. One working, one done and saying so.

When a session needs my permission, its robot hops, waves, and says what it wants to do in a speech bubble. It keeps at it until I answer. When a session finishes, the robot tells me what it did.

When Claude Code hands part of a job to a helper agent, a small drone flies out of the robot and reports back.

Quiet sessions look quiet. After a few minutes with nothing happening, a robot dozes off. After an hour it wanders away, and it comes back when the session picks up again.

Each new session gets a freshly generated robot and keeps that look until it ends. When things are slow, the robots wave, nap, play catch, and make small talk.

How it works

It listens to Claude Code’s hooks, the events Claude Code announces as it works. A small helper writes each session’s state to a file on my Mac, and the app reads that file. The session data never leaves the machine.

How it got made

I built it in a few days with Claude Code. I knew the problem. What took a couple of rounds was working out which situations the robots needed to communicate. It is a native Mac app, signed and notarized, and it updates itself.

Before anyone at work installed it, I cleared it with our head of technology and with IT. Twenty to thirty coworkers run it now. A few came to find me when they learned who had made it. Fun, useful little things do not usually just show up at our company.

What I took from it

As coding agents do more of the work, someone has to design how you keep track of them: what they are doing, and what they need from you. I learned that by having the problem.

What is next

This was a quick proof of concept, and I always planned a second version. Then OpenAI launched Dots, agents with little avatars of their own, and I like theirs better than mine. That stings a little. It also makes me want to start version two yesterday.

On the list: click a robot to jump to its terminal, a sound when one needs me, and robots from the same project standing together. In the meantime, the first version works fine. It sits on my desktop every day.

All experiments

Image