CodeCat

You give an agent a long task, close the lid, walk away. Two things go wrong: the agent stops to ask a permission question and sits idle for an hour, or the Mac falls asleep and cuts the session mid-work.

The existing tools each solve half of it. caffeinate and Amphetamine keep the Mac awake but have no idea whether an agent is still working. Claude Code's own notification hook fires once and forgets. RunCat reacts to CPU, and a waiting agent uses none.

I wanted one thing on screen that knows the state of every session and reacts to it: keep the Mac awake while agents work, and get my attention the moment one needs me.

CodeCat is a native macOS menu-bar utility. It subscribes to Claude Code's hooks, keeps a live model of every session — CLI and desktop app — and drives three things from it: an IOKit power assertion, a menu-bar item, and a pixel cat.

The cat is the interface. It sleeps when nothing is running, works with a badge showing the session count, waves a paw with an orange badge when an agent is waiting for you, and stretches and lies down when everything is done. It lives either as a floating always-on-top window or as an "island" — a black plate that covers the MacBook notch, cat on the left, session counter on the right.

Clicking a session takes you to it. In Terminal and iTerm2 that is the exact tab, found by its tty. In the desktop Claude app it is the exact chat: the app keeps a record per session and registers a claude://code/continue link, and since 0.4.0 the row opens that chat instead of just bringing the app forward. A "while you were away" panel collects everything that happened between screen lock and unlock. Closed-lid mode uses pmset disablesleep behind a sudoers rule scoped to exactly two commands, with a LaunchDaemon that clears the flag if CodeCat ever dies.

In 0.4.0 the cat also learned to wear any pet drawn for the Codex pet format — a folder with a pet.json and an 8 × 9 sprite sheet, read from ~/.codex/pets or from CodeCat's own Pets folder. The format is someone else's, which set the rules: pets are read where they are, never copied or redistributed, and a broken folder is skipped and named once in the log. They often ship upscaled, so CodeCat finds the sheet's native pixel pitch and draws it as crisply as the built-in packs whenever the native drawing fits the cat's canvas. Sessions are still Claude Code only: wearing a Codex pet doesn't make CodeCat track Codex.

0.4.0 counted sessions. The next release, built on 18 September, draws one dot per session in its own colour — up to four, then a count — and lets the worst state tint the glow behind the cat. Hover, and each row says the task in your words, the step the agent is on and how far along its own plan it is, read from the TodoWrite list it keeps in the transcript. A finished turn hands back what it made — a local server, a pull request, a Figma link — as chips you can click or drag.

I wrote the spec at 23:10 on 28 August. The agents wrote their first code 17 minutes later. Working installed build on day two, signed and notarized on day five: 151 commits, roughly a third of them post-review fixes. I had every mascot state rendered to PNG and checked each one by eye. That is how I found that all the animations were attached to an EmptyView(): the cat had no body.

No analytics, no network calls, no dependencies. The download is a 2 MB notarized disk image.