AgentWatch

I run a lot of coding agents at once. Claude Code in a terminal, another one in Xcode, a Codex session somewhere behind that, and an SO-ADK pipeline fanning out to twenty sub-agents while I work on something else. The problem isn't starting them. It's that once a terminal window is buried, an agent that finished twenty minutes ago and an agent that's waiting for my approval look exactly the same: invisible.

AgentWatch puts all of it in the menu bar. One icon, every running session behind it, and the cost adding up while I work.

What it is

A row per session: project, model, how long it's been going, what it has spent. A coloured dot on the left carries the status, and that dot is the only colour in the entire app. Claude quota comes from the same OAuth endpoint the CLI authenticates against, so those percentages are real. Codex publishes no quota API at all. Its numbers get parsed out of local rollout logs, and the app says so on screen rather than presenting a guess as a fact.

It started as a fork of so-agentbar by sotthang, MIT licensed. The session-monitoring core underneath is largely his work and it's good work. What I changed is everything above it.

How I think about it

• Colour means one thing. The interface is greyscale on purpose, so anything coloured on screen is an agent that wants something from me. Status badges, source icons and meters all lost their hues to make that true.
• Measure, don't eyeball. Layout decisions here were made by rendering views offscreen and counting pixels, because "looks about right" kept being wrong after I swapped in bundled fonts with wider metrics.
• A fork inherits more than code. Signing identity, update keys, feed URLs and release tags all came along, all pointed at someone else, and each failed differently.

The article below is about the rebuild and about the release pipeline, which is the part I actually learned something from.