Replies and quota alerts on your Mac
When an agent finishes a turn or a quota crosses a line, the HUD tells you on the display you are looking at. On the Mac, alerts appear only in the HUD; Agent HUD never posts to Notification Centre.
Updated 2026-10-26
A new reply
A finished turn arrives as a New reply card: the agent and model, the session's title, the opening lines of the reply, when it finished and how long the turn ran. Return to session takes you to its thread, tab or terminal pane in one click, and View token usage opens what the session used.
Replies come from Claude Code, Codex, DeepSeek Harness, Antigravity, Cursor, Grok CLI, GitHub Copilot CLI, OpenCode, Kimi, Pi, CodeBuddy, Qwen Code and Devin CLI. One click back works for Claude Code, Codex, Antigravity, Grok Bot, OpenCode and Pi, from a reply or from the session's page.
A turn that finished before Agent HUD started is not replayed.
Quota alerts
- 90% used: a quota crosses 90% of its window.
- Used up: nothing is left until the reset.
- Runs out before the reset: at the recent burn rate, the window will be empty before it resets.
- Reset: the window starts again, early resets included. The alert names any other window of the account that is still used up.
- Usage resets added: a Codex account received resets it can spend.
Once, and only for a change
Each event comes once for a window and its cycle. The first reading after Agent HUD starts sets a baseline, so a window already at 95% raises nothing until it changes. A reading older than 30 minutes stays on screen but raises nothing. Between alerts, the glow carries the state: green under 70% used, yellow from 70%, red from 90%.
Where and when they appear
- On the display under your pointer. In Notch mode they open at the notch; in Simple and Dynamic they come out of the sender's logo. In Notch and Dynamic the glow flashes in the alert's colour.
- One at a time, in the order they came. A reply or a quota alert closes 4 seconds after it shows, or after the pointer leaves it, never while you hover it.
- A request stays until you answer it or it goes back to the agent. While one is on screen, replies wait their turn and quota alerts are not shown.
- Silently. Agent HUD plays no sound and posts nothing to Notification Centre.
What Claude Code and Codex do on their own
Claude Code sends its own desktop notification when it finishes or waits for a permission and you seem to be away, by default in Ghostty, Kitty and iTerm2. Other terminals can ring the terminal bell instead, or run a Notification hook.
Codex can hand each notification to a command you name, as a JSON payload, and its terminal interface can post notifications of its own.
Each of those notices comes from one agent. Agent HUD gathers the replies of every agent it follows in one place, and adds the quota alerts.
Claude Code: terminal bell and notificationsCodex: configuration reference
On iPhone and Apple Watch
With Agent HUD Remote, alerts reach your iPhone and Apple Watch, and running sessions show as Live Activities on the Lock Screen and in the Dynamic Island.
Questions
Does Agent HUD put anything in Notification Centre on the Mac?
No. On the Mac every alert appears in the HUD. Alerts on iPhone and Apple Watch are part of Agent HUD Remote.
Why was there no alert for a window already at 95%?
The first reading after Agent HUD starts is a baseline. An alert comes when a window crosses a line while Agent HUD is running.
Will the same alert come twice?
No. Each event comes once for a window and its cycle, so a window that stays above 90% stays quiet.
Why does a build from the GitHub source show fewer replies?
It does not install client hooks yet, so requests and the replies that arrive through hooks stay in each agent's own window. The release installs them.