The app, one screen at a time
Most of the time, it's quiet
Agents run, you work, and the notch stays out of the way — until one of them stops and needs you. Here is every screen it shows, paired with the moment it saves.
In the notch
Then one needs you
Approvals
The command, and why it asked
An agent stops to run something risky. You see the exact command and the reason before you Deny or Approve — from the notch, without switching windows.
Questions
It asks in words, with options
When an agent needs a decision it can't make alone, the question and its choices surface — so it waits on an answer instead of guessing.
The fleet
Run six. Watch none.
Every agent with its branch, its context burn-down and how long since it moved. Only one thing is ever coloured: the lamp on the session that wants you. Everything else is quiet on purpose.
Every terminal
Ghostty, Cursor, Codex, Kimi — one place
It reads the agents themselves — Claude Code, Cursor, Codex, Kimi, OpenCode, Antigravity and Pi — across worktrees and terminals, and flags the one burning down its context before it stalls. The Cowork sessions in your Claude app land in the same list, and so do local models running in LM Studio.
Plan limits
Enough left to finish?
One track per window a provider actually meters — Claude's five hours and its week, Codex's week, Cursor's month, the OpenCode Go plan. Burn one faster than it refills and it stops counting down to the reset and starts counting down to empty.
