A tiny menubar app that reads your local Claude Code transcripts directly — live percentages that match claude.ai, refreshed within seconds, with no API calls.
make app CONFIG=release
Three meters tell you whether you can keep working, switch profiles, or need to slow down. Click for the full picture.
Today, weekly all-models, and Sonnet-only — percentage and bar, color-coded the moment a threshold breaks.
Hold any number of API keys + plans. Switch with a click; the menubar accent shifts so you always know which you're spending.
Plan-aware projection: a token usage-pace forecast for subscription plans (Max, Pro, Team), a dollar burn-rate for pay-as-you-go API keys — so you spot a problem on day six, not day twenty-eight.
The bits Anthropic only exposes in the dashboard are one keyboard shortcut away — opened in your default browser, in the right account.
Claude Helper polls the same endpoints your terminal hits and renders them where you'll actually look.
Paste an API key. Helper stores it in the macOS Keychain — never on disk in plaintext, never on a server.
A lightweight background daemon reads your local Claude Code ~/.claude/projects/ transcripts every 2 minutes. Byte-offset caching means incremental runs take ~100ms. The menubar updates within 500ms of any new token write via FSEvents.
Yellow = slow down. Red = switch profiles or raise the limit. The popover opens to the action you need.
Each profile bundles an API key, a claude.ai web session, and the plan tier. The active profile colors the menubar arc so you always know which account is being charged for the next request.
On a subscription, Helper projects your month-to-date token consumption against your all-models cap — no fictional dollar figure, because a Max plan has no per-token cost. On a pay-as-you-go API key it draws the dollar burn-rate instead. Either way: cross the line on day six, see red on day seven — not when the limit hits.
Set a percentage — 80% by default — and Helper fires a single native macOS notification when you cross it. No menubar bouncing, no dock badges, no email.
No backend, no analytics, no telemetry endpoints. API keys live in the macOS Keychain. Helper talks to api.anthropic.com directly — and to nothing else.
Stored alongside Safari passwords. Sandbox-restricted to Helper's bundle ID.
MIT licensed. Build it yourself; verify the network calls.
No sign-up. No license server. Quit and uninstall in five seconds.
SwiftUI + AppKit. ~6 MB on disk. ~30 MB of RAM. No Electron.
Claude Helper isn't on Homebrew. The prebuilt bundle is ad-hoc signed because the maintainer doesn't run an Apple Developer account, so Gatekeeper will warn on first open. Build from source for the strongest trust path — or verify the SHA-256 below.
macOS 15+, Xcode 16+, Swift 6.0, GNU make. The binary never touches the internet, so Gatekeeper leaves it alone.
$ git clone https://github.com/peguesj/claude-code-macos-helper.git $ cd claude-code-macos-helper $ make app CONFIG=release $ make install # copies to /Applications and launches
Ad-hoc signed universal .app packaged as a zip. Verify the SHA-256 before opening.
c025dfdb1b29a2e8d9bc93682bf0d7997e504728af41d03350635b5884c48a0d
$ shasum -a 256 ClaudeHelper-v0.1.1.zip # must match the hash above before you unzip $ unzip ClaudeHelper-v0.1.1.zip $ mv ClaudeHelper.app /Applications/ $ xattr -dr com.apple.quarantine /Applications/ClaudeHelper.app $ open /Applications/ClaudeHelper.app
Ad-hoc signed ≠ verified. Anyone who can swap the zip on the release page can ship a different binary. The SHA above is published from this site (separate channel from the binary CDN). If you can't run shasum, build from source instead.
make app CONFIG=release — you trust your own toolchain, no Gatekeeper friction. If someone wants to sponsor a Developer ID, open an issue.