I Forked an Open-Source Mac Utility. Here's What I Changed, What I Kept, and When You Should Use the Original

Most AI coding tools now meter you in windows. A five-hour session here, a weekly cap there, a monthly billing cycle somewhere else, and every provider shows it on a different page. When you use more than one of them, "how much do I have left before this resets?" turns into three browser tabs and a guess.

That question is how I ended up forking an open-source Mac app and selling the result. Since that combination makes people suspicious (fairly), this post is the full picture: what the original does, what I changed, what I kept, and who should skip my version entirely.

The original: CodexBar

CodexBar is a macOS 14+ menu bar app by Peter Steinberger, released under the MIT license. Its tagline is "Every AI coding limit, in your menu bar," and it delivers on that. At the time of writing it's on v0.84.0 with about 23k GitHub stars, and its README lists over 100 providers.

A few things it does that I think are well designed:

• It reuses sessions you already have. Depending on the provider, it reads from OAuth, a device flow, an API key, browser cookies, local CLI files, or a CLI fallback, so you don't hand it a password.
• Per-window reset countdowns. Session, weekly and monthly windows each get their own countdown, which is what you actually need before you start a long agent run.
• It goes well past the menu bar. It ships a bundled codexbar CLI (with codexbar cost for local Codex and Claude cost scans), WidgetKit widgets, provider status polling with incident badges, a Linux desktop app, and a 23-language UI.
• It's documented. Separate docs cover architecture, the refresh loop, provider authoring, and exactly which macOS permissions it asks for and why.

If you have never used a usage meter in your menu bar, install CodexBar first. It's free and very good.

Why a fork and not a pull request

CodexBar describes itself as a "tiny" menu bar app with "minimal UI." The things I wanted to build don't fit that description: a full dashboard window, alerts about what coding agents are doing (not just quota), an embedded pixel-art view of agent sessions, and a system load panel. Proposing all of that upstream would mean asking a deliberately small project to become a bigger one. A separate app was the cleaner option.

The MIT license makes that legal, but it doesn't make it automatically fair. The bar I set was that the fork had to add real work on top, keep the license and attribution intact, and say plainly where it came from. TokenBar's trust page states that its repository history descends from CodexBar and includes later contributor work as well as my changes.

What I changed

These are TokenBar features you can check on the public site today:

A dashboard window, not only a menu. TokenBar adds a "Your AI, at a glance" window with a spending overview: recorded usage value for today, cost history over the past 30 days, known monthly plan prices, the number of tracked providers, and a 30-day chart for your top model. One decision I'd make again: the dashboard keeps estimated usage value and subscription prices as separate numbers, and says on screen that usage value may include token-price estimates and that your provider's invoice confirms actual charges. Mixing the two into one "you spent X" figure would look cleaner and be wrong.

Worth knowing: CodexBar has its own Usage & Spend view with model breakdowns and daily trends, so this is a difference in presentation more than in raw data.

Pixel Agents inside the app. Pixel Agents is a separate MIT-licensed open-source project that draws your coding sessions as pixel-art characters in a small office. TokenBar has a panel that connects to it. TokenBar doesn't bundle Pixel Agents. You point it at your own local build (it needs Node.js 20 or newer), so what's supported depends on the version you installed, and the Pixel Agents project keeps ownership of its own code.

Agent notifications. On supported integrations, you can get configurable alerts when a task finishes, an agent asks for input or an approval, or a session stalls, with cooldown and timeout controls. CodexBar has notifications too, but its README describes them around quota: session quota alerts, reset warnings and weekly-reset confetti.

System load. A panel showing CPU, memory, energy and heat. When several agents and a local build run at once, the machine is part of the picture.

Appearance. Light or dark appearance, a customizable dashboard, and a choice of Pixel Agents office background.

What I kept

Looking inside the current TokenBar app bundle (version 0.38.1) shows how much of the upstream shape is still there:

• A bundled command-line helper, tokenbar, with JSON output, a cost command, and config validate and config dump, close to how CodexBar's CLI is organized.
• A WidgetKit widget extension.
• Provider resources for 23 services, including Claude, Codex, Cursor, Gemini, Copilot, OpenRouter, Vertex AI, Windsurf and JetBrains. The website focuses on the first five.
• Menu bar meters with reset times for each provider.

Side by side

Checked against the CodexBar README, tokenbar.site, and the TokenBar 0.38.1 app bundle on November 12, 2026.

Neither app gives you more quota. Both only show you what the providers report.

Who should just use CodexBar

Honestly, most people reading this:

• You want limits and reset times in the menu bar, and nothing else.
• You need a provider TokenBar doesn't cover. CodexBar's list is far longer.
• You're on Linux, or you want a CLI with a large community of integrations around it (SketchyBar, tmux, Stream Deck and more are listed in its README).
• You like reading the source of anything that touches your provider sessions.
• You'd rather not pay for this category of tool. That's a fine position.

TokenBar makes sense if you want the dashboard window, agent alerts and Pixel Agents in one app, and you're happy paying for that.

How the pricing works

There are two plans, both per Mac. Basic is $5 a month for core TokenBar access. Pro is $15 once, a lifetime license that adds the premium dashboard tools, provider mix and richer insights. You download the app first and choose a plan inside it; there's no checkout on the website.

What you're paying for is the work described above, a maintained build from tokenbar.site, and email support. What you're not paying for is the open-source core, which stays free in CodexBar for anyone who wants it.

What I'd tell anyone thinking about a paid fork

1. Read the license, then do more than it requires. MIT only asks you to keep the notice. Name the original author where people will actually see it.
2. Write down what you added in a form people can check. If you can't list it, you don't have a product yet.
3. Assume the upstream will keep improving. CodexBar now has spend views and burndown charts, so a dashboard alone is not a reason to pay. Your differentiation has to survive that.
4. Recommend the original when it fits. Nobody trusts a comparison that only has one winner.

Disclosure: I build TokenBar at Aura Technologies. It's built on CodexBar by Peter Steinberger (MIT), and its agent office comes from the Pixel Agents project (MIT). If you want to look at it, it's at tokenbar.site.

A question for people who've built these

CodexBar's provider list shows how many ways there are to get a usage number out of a service: official APIs, OAuth, device flow, browser cookies, CLI output, local log files. If you've built anything that reads usage or quota data, which of those sources has held up best for you when the provider changes something, and which broke first?