Leeway

How much of your Claude Code and Codex limits you've used — right in the menu bar.

 

The menu bar shows something like this:

63% · 2:14

63% of your 5-hour window is used, and it resets in 2 hours 14 minutes. Click it to see the rest: the 7-day window, when each one resets, and which model your next session will start on.

Leeway follows Claude, Codex, or both at once. In the menu bar, a dot in front of each number tells you whose it is — orange for Claude, grey for Codex.

What you can do

• Watch your limits — the 5-hour session and the 7-day window, for Claude Code and Codex.
• Start a 5-hour window when it suits you. The window starts with your first request, so if you start it at 9:00, it resets at 14:00 instead of in the middle of your afternoon.
• Pick the model and effort for your next session without opening a terminal.
• Make the menu bar yours — show the 5-hour or the 7-day window, percent used or percent left, with or without the time until reset.

What you need

• macOS 15 or later.
• For Claude: Claude Code, signed in with a Claude subscription.
• For Codex: the Codex CLI, signed in with your ChatGPT account (codex login).

With an API key, neither tool reports subscription limits, so Leeway has nothing to show.

Where the numbers come from

Leeway reads only what Claude Code and Codex already tell you. It doesn't scrape websites, and it never reads your passwords or tokens itself — the two tools sign in on their own.

Claude. Every 30 seconds to five minutes, Leeway runs Claude Code's own /usage report. It costs nothing and needs no setup.

For live numbers, you can also install the status line hook in Settings → Claude Code → Status line. Claude Code then hands Leeway fresh numbers with every reply, along with the model that's answering. Your existing status line keeps working: Leeway takes a copy of the data and passes it on. Before the first change, Leeway saves a copy of your settings as ~/.claude/settings.json.leeway-backup, and removing the hook puts your old status line back.

Codex. Leeway asks the Codex CLI on your Mac for your limits (account/rateLimits/read). This doesn't start a conversation or spend anything.

Starting a 5-hour window

Neither Claude nor Codex runs the 5-hour window on a fixed clock: it starts with your first request after the last reset. Start 5-hour window sends that first request for you.

• Claude: a tiny request to Haiku. It costs next to nothing.
• Codex: there is no free way to do this, so the button spends a couple of thousand tokens of the window it opens. The request ignores your config.toml (your model, effort, MCP servers and hooks are for your work, not for this) and stays out of your session history.

Model and effort

The Model menu sets what your next session starts with — the same thing /model and /effort change. A session that's already running keeps its model.

Each model in the menu opens a list of effort levels, so you pick both in one click. Default removes your choice and lets the tool decide. Max effort isn't offered: it only ever lasts one session.

• Claude: Leeway writes the model key in ~/.claude/settings.json and the effort level under modelSettings, where /effort keeps it too. Models are saved by family name, so opus always means the newest Opus.
• Codex: Leeway asks the Codex CLI to make the change, so config.toml keeps your comments and formatting.

If your shell profile sets ANTHROPIC_MODEL or CLAUDE_CODE_EFFORT_LEVEL, that wins over any setting, and the panel tells you so. Leeway reads your shell profile at launch, so restart it after you change the profile.

Good to know

• — means there's no fresh number yet — for example, right after a window resets.
• Running several sessions at once is fine: they all share one limit.
• If a refresh fails, the last number stays on screen with its age and the error.
• Codex limits for individual models aren't shown, only the account's 5-hour and 7-day windows.

Building from source

Xcode 26. The version comes from the git tag and the build number from the commit count, so build from a checkout that has its tags. SwiftLint runs as the first build phase.

To test reading Codex limits, the model menu and the process handling — without signing in or touching the network:

swiftc -module-cache-path /tmp/leeway-swift-cache \
  Leeway/Support/{ChildProcess,LoginShell,MenuBarFormat,ModelMenu,UsageSnapshot,CodexCommand,CodexAppServer,CodexUsageCommand,CodexModelCommand}.swift \
  scripts/CodexUsageChecks.swift -o /tmp/leeway-codex-checks
/tmp/leeway-codex-checks

Add --live to test against the Codex CLI you have installed and your real account.

License

MIT — see LICENSE.

Leeway is an independent project, not affiliated with or endorsed by Anthropic or OpenAI. Claude and Claude Code are trademarks of Anthropic; Codex is a trademark of OpenAI.