Pastor de Agentes

I work with several Claude Code and Codex agents at once, on different projects. I didn’t know which one had finished, which one was stuck and which one had been waiting for my permission for ten minutes.

Or how much of my limit was left before I got stranded halfway through a task.

• All of them in the Mac’s menu bar. A sheep in the menu bar with every agent grouped by project. Click one and you jump to its session.
• Instant alerts. When an agent finishes, gets stuck or asks for permission. And you can grant it from the alert itself.
• Stray sheep. If an agent chains five errors in a row, it’s in a loop. If it’s been burning tokens for 45 minutes without delivering anything, it warns too.
• Limits in sight. How much is left on Claude and Codex, with a warning at 80% and a forecast of when it runs out.
• The pen. The whole week in a calendar, with each project’s work and whatever was left unsaved.
• The accounts. Hours per project, exportable, to account for what was done.
• Memory. A search across every session and a day-by-day diary.
• From my phone. If I want, alerts reach me on Telegram, with buttons to grant permission.

I wrote it down first: what it had to watch, in what order and what not. The first version came out in a single day, and then I extended it in phases, each with its own design and tests.

To test it without waiting for an agent to really go astray, I built a player that replays old sessions.

Now I always know what each agent is doing, which one needs me and how much I have left.

It’s for me. It isn’t for sale.

Swift and SwiftUI for macOS. SQLite to keep the history on the Mac. It reads Claude Code and Codex sessions straight from their files.

• Pincho · The screenshot app the Mac was missing.
• Evolve Performance · Gym apps, in real use.
• My TV · My Fire TV’s system, rebuilt to explore films.