RobotBar

A native macOS menu bar app: a little animated robot that comes alive when Claude Code is generating, plus optional ChatGPT app activity detection.

Requirements

• macOS 14 (Sonoma) or later
• Xcode 15+ (or the Swift toolchain via xcode-select --install)

Running it

Build the app bundle:

cd RobotBar
zsh Scripts/build-app.sh
open .build/RobotBar.app

That creates .build/RobotBar.app, a double-clickable menu bar app. It does not show in the Dock; look near the clock/control center area for RB and the small robot face.

Fastest way to run without packaging:

cd RobotBar
swift run

To develop/debug in Xcode: open the RobotBar folder. Xcode will recognize Package.swift and let you build/run/breakpoint it like a normal app target. No .xcodeproj needed.

Options

Open the RobotBar menu and enable ChatGPT app activity if you want the robot to light up when the ChatGPT macOS app is frontmost. This setting is saved in UserDefaults.

Use Choose Folder under Claude Logs to select your ~/.claude folder, or a Claude log subfolder such as sessions or projects. This is optional for local development, but it is the right model for a Mac App Store sandboxed build because the user explicitly grants read access.

ChatGPT activity detection is presence-based only. It does not include token counts because the ChatGPT app does not write Claude Code-style local usage JSONL files.

Mac App Store prep

See AppStore/APP_STORE_CHECKLIST.md for the remaining submission steps. RobotBar now includes AppStore/RobotBar.entitlements with App Sandbox and read-only user-selected file access. A final App Store build still needs a real Xcode macOS App target, app icon, Apple Developer bundle identifier, signing, and archive/upload through Xcode.

How it works

• FSEventsWatcher.swift - a thin wrapper around Core Services' FSEventStream, recursively watching the selected Claude logs folder and reporting changed file paths the moment Claude Code writes to a session log. This is event-driven, not polling.
• ActivityMonitor.swift - on each Claude Code change, incrementally reads only the newly-appended bytes of the changed .jsonl file, extracts each assistant turn's message.usage object, de-dupes by message.id, and aggregates into "today" and the current rolling 5-hour window. It also observes the frontmost macOS application so the optional ChatGPT activity mode can light up the robot when ChatGPT is active.
• RobotIconView.swift - a hand-drawn menu bar robot with an RB label so it is easier to find in a crowded menu bar.
• MenuBarExtra(...).menuBarExtraStyle(.window) - lets the dropdown be an arbitrary SwiftUI view instead of a plain native menu list.

Where the data comes from

Claude Code writes local JSONL transcripts under ~/.claude. Depending on the Claude Code version and configuration, those logs may be under subfolders such as projects or sessions.

Every assistant turn is a JSON line containing a message.usage object with the token breakdown, and a message.model field. This is entirely local: no API key, no network call, nothing leaves your machine.

Important limitation: ChatGPT and Claude Desktop token counts are not covered

Claude Code writes local token logs; the regular ChatGPT app, Claude Desktop chat window, and claude.ai web chat do not expose the same local usage files. RobotBar can light up for ChatGPT app focus, but token counts remain based on Claude Code logs.

Claude Desktop's Cowork/agent mode is a partial exception: it writes its own audit log to ~/Library/Application Support/Claude/local-agent-mode-sessions/, which could be added as a second watched path later.

Ideas for next passes

• Add ~/Library/Application Support/Claude/local-agent-mode-sessions/ as a second watched directory to also catch Cowork sessions.
• Add similar frontmost-app toggles for other AI apps.
• Read the Claude Code OAuth token from the macOS Keychain (service Claude Code-credentials) to optionally show official plan/rate-limit percentage from the API, not just locally-derived totals.
• Persist seenMessageIDs / offsets to disk so counts survive an app restart.
• Support Codex/other CLI tools' local logs the way ai-token-monitor does, if you want this to go beyond Claude specifically later.