AgentHub

AgentHub is a macOS menu bar panel that shows how much of your coding-agent subscriptions you have used.

It reads usage for four providers and shows them in one place:

Hover the menu bar icon to see them, click to pin the panel, or press ⇧⌘T from anywhere — the shortcut is rebindable in Settings. Usage refreshes every fifteen minutes, and the pinned panel has a refresh button. The − and + buttons zoom the whole panel, from 80% to 250%, and the size is remembered.

A provider that reports nothing says why instead of leaving a gap.

Settings chooses which providers appear. A hidden one is not merely left out of the panel — it is never contacted, so hiding Codex stops spawning a subprocess and hiding Cursor stops calling cursor.com.

AgentHub used to manage agent sessions — listing them, approving permission requests, handing context between agents. That is gone. Only Cursor exposed live desktop sessions; Claude Desktop cannot be observed at all, and Codex Desktop threads are not reachable from an app server we can start. The reasoning is in docs/superpowers/specs/2026-08-13-agenthub-quota-menubar-design.md.

xcodegen generate
xcodebuild -project AgentHub.xcodeproj -scheme AgentHubApp \
  -configuration Debug -destination 'platform=macOS' \
  -derivedDataPath .build/xcode CODE_SIGNING_ALLOWED=NO build
open .build/xcode/Build/Products/Debug/AgentHubApp.app

zsh scripts/check.sh runs the whole gate: package tests, the app test bundle, and the source rules below.

AgentHub has no Dock icon and no window — it lives in the menu bar. Nothing starts it for you, so add it to Login Items if you want it always there.

On first launch it removes what older versions installed: the com.agenthub.daemon LaunchAgent, AgentHub's Claude hooks and status-line wrapper, and AgentHub's Cursor hooks. Hooks belonging to other tools are left alone, and both files are backed up first.

Cursor is opt-in: until you authorise it in Settings, AgentHub does not read the session token Cursor stores on this Mac.

Claude works as long as Claude Code is signed in. Claude Code keeps its credential in ~/.claude/.credentials.json or in the login Keychain, and a Mac can hold both — one of them a dead leftover from an older sign-in — so AgentHub reads whichever has the later expiry rather than preferring either.

Claude's access token lasts eight hours and Claude Code only renews it when it makes a request, so a Mac that has not run it since this morning holds an expired one. AgentHub then performs the same refresh Claude Code performs and writes the result back to the same store, so a rotated refresh token does not log the CLI out. This is the one credential AgentHub writes, and it writes it only where Claude Code already keeps it. If the refresh token is itself dead, the panel says so and claude auth login is the fix.

Codex is the only provider AgentHub shells out to. An app launched from Finder inherits launchd's environment, which is nearly empty — no PATH pointing at per-user CLI installs, and no proxy variables. That broke Codex twice: once because codex could not be found, once because it started but could not reach the network.

So AgentHub asks your login shell what it exports ($SHELL -ilc, once per launch) and hands that to codex, giving it the same environment the command would have in your terminal. Nothing is stored, and the shell is only ever asked to print its environment.

URLSession reads the system proxy configuration by itself, so the other three providers never needed this.

Every provider token is read at the moment of the request, kept in memory for that request, and never written to disk, a log, or the panel. scripts/check.sh enforces this: token literals may appear only in the three reader files that need them, and never in the app.

AgentHub never installs software or shells out to a package manager to obtain usage.