tokn.watch vs ccusage: menu bar vs CLI for Claude Code

Use ccusage if you want to know what your Claude Code sessions cost in tokens and dollars, broken down by day, project, and model. Use tokn.watch if you want to know how much of your plan is left right now, across Claude and ChatGPT, without opening a command line. They measure different things, and running both is perfectly reasonable.

The 30-second verdict

We write the comparison pages on this site, which is a good reason to read them sceptically. So here is the summary first, including the rows where we lose.

The two decisive rows are core number and data source. Everything else in this article follows from those.

What ccusage does well

ccusage is a fast, free, open-source reporting tool with more than 18,000 stars on GitHub. You do not install it globally: you run npx ccusage or bunx ccusage and it reads the log files your coding agents already write. There is no account, no API key, and nothing uploaded.

It covers far more than Claude. The project describes itself as a tracker for Claude Code, Codex, OpenCode, Amp, Droid, Gemini CLI, GitHub Copilot CLI and a dozen more agents. If your week is spread across several coding agents, ccusage is one of the very few tools that adds them up in a single report.

The report shapes are genuinely useful. The documented commands are daily, weekly, monthly, session, blocks, and statusline. So the questions it answers well are "what did today cost", "which project ate the month", and "which model am I actually using". It separates cache creation from cache read tokens, supports custom pricing overrides, groups by your own timezone, exports JSON for scripting, and has an offline mode with pre-cached pricing.

The blocks command understands Claude's 5-hour window. Per the ccusage documentation, a block starts with your first message and lasts exactly 5 hours. Running ccusage blocks --active shows only the window you are in, with time remaining, your current token burn rate, and a projection of where the window ends up if you keep going at that rate. For a heavy Claude Code session that is a real early-warning signal.

It is not tied to a Mac. It is installed through Node or Bun package runners, so it works on whatever machine you code on. tokn.watch is macOS only. If you work on Linux or Windows, this comparison is already over.

What changed in ccusage recently

Worth knowing if you are following older tutorials: the ccusage docs state plainly that the blocks --live monitor was removed in version 18.0.0. The replacement is ccusage statusline, which you wire into Claude Code's own status line through ~/.claude/settings.json. It prints one line with the active model, session cost, today's cost, the current block cost with time left, burn rate, and context use. The docs mark it beta. The current published version at the time of writing is 20.0.22.

The one thing a local log reader cannot see

This is the structural difference, and it is not a criticism of ccusage's engineering. ccusage reads files like ~/.claude/projects/ on the machine you run it on. That is the honest, privacy-preserving way to do it, and it is why the tool needs no credentials at all.

But your subscription limit is not a property of one machine. It is a property of your account, and it is shared across Claude Code, the Claude apps, and claude.ai. Anthropic makes exactly this point about its own built-in command. The Claude Code cost documentation says the /usage breakdown figures are:

approximate and computed from local session history on this machine, so usage from other devices or claude.ai is not included. Anthropic, Claude Code documentation

Any tool that works from local logs inherits that boundary. If you spent the morning on claude.ai in a browser, coded on a laptop in the afternoon, and then ran a report on your desktop, a local log reader can only ever show you the slice of the day it has files for. Your weekly cap counted all of it.

That gap is the reason tokn.watch exists: it asks your account what is left rather than adding up what one machine did. Anthropic's own documentation describes a five-hour session limit and a weekly usage limit on paid plans, and the usage limit best practices page points you at Settings then Usage to see where you stand. A menu bar meter is that page, without the trip to the browser. We go through the underlying windows in detail in our guide to how Claude's usage limits work.

Dollars spent versus percentage left

These two numbers feel interchangeable and are not. A cost figure is backward-looking and derived from token counts multiplied by a price list. A remaining percentage is forward-looking and comes from the provider's own accounting of your allowance.

For subscribers, the cost figure is doubly indirect. Anthropic's documentation notes that on Max and Pro the session cost is not relevant for billing, because that usage is included in the subscription, and that the dollar figure is computed locally at list price as an estimate. It is a proxy for effort, and a good one. It is not a fuel gauge.

What tokn.watch does differently

tokn.watch is deliberately narrow. It shows the remaining session and weekly percentages plus reset countdowns for Claude and consumer ChatGPT, in one menu bar item, refreshing on its own between every minute and every hour. It is free, has no paid tier, no licence key, and no account to create. It runs on macOS 13 and later, is Apple-signed & notarized as a direct download, and is also on the Mac App Store.

Three differences matter most against a CLI tool:

• It is passive. You do not decide to check. The number is already on screen when you are about to start something expensive. That is the entire design argument for a menu bar meter over a command you have to remember to run.
• It covers chat, not just coding. Most of the surprise limit hits we hear about come from a mix: a long claude.ai conversation in the morning and coding in the afternoon, sharing one pool. ccusage sees only the coding half.
• It covers ChatGPT too. ccusage reads Codex, the coding agent. It does not read ordinary chatgpt.com use, because that leaves no file on your Mac to read.

And three things it does not do, which ccusage does: no per-project cost, no history or trend, and no dollar figures at all. If those are your questions, tokn.watch is the wrong tool and we would rather say so here than have you find out after installing.

Where ccusage is genuinely the better choice

Comparison pages written by one of the two products usually bury this section, so here it is in full.

• You do not use a Mac. tokn.watch is macOS only. ccusage is not.
• You bill clients for AI work. Per-project and per-session cost with JSON export is exactly what an invoice needs.
• You use several coding agents. Claude Code plus Codex plus Gemini CLI in one report is something no menu bar meter offers.
• You want history. Daily, weekly, monthly and session views reveal patterns that a current percentage cannot.
• You want it inside the coding session. The statusline integration puts the numbers where you are already looking.
• You want to script it. JSON output makes ccusage a building block. tokn.watch is an app, not a data source.

Where tokn.watch is the better choice

• You use Claude outside the coding agent. Chat in the browser counts against the same pool and leaves no local trace.
• You use more than one machine. Local logs are per-machine; your limit is not.
• You also pay for ChatGPT. Two subscriptions, one menu bar item.
• You want the warning without asking for it. A visible percentage beats a command you run after the lockout.
• You do not want to think about package runners. Install, sign in, done.

Running both is the honest recommendation

There is no conflict. Neither tool touches the other's data, neither needs an account with the other, and they read different things: one reads files on disk, the other asks your provider. The pattern we see most often among heavy users is a meter kept visible all day for the ambient signal, and a CLI report run on Friday for the breakdown.

If you want to set that up, the shortest path is: install the menu bar app and connect it once using the Claude connection steps on our homepage, then run npx ccusage daily whenever you want the cost view. Neither step affects the other. We cover the other combinations, including the built-in Settings page and /usage, in how to check your Claude usage on a Mac.

So which one should you install?

1. Claude Code is most of your usage and you want cost analysis: ccusage. Start with npx ccusage daily.
2. You keep getting surprised by limits: tokn.watch. A backward-looking cost report cannot warn you.
3. You pay for Claude and ChatGPT: tokn.watch, because ccusage does not read consumer ChatGPT.
4. You code on Linux or Windows: ccusage, and it is not a close call.
5. You want both angles: both. They cost nothing together.

If you are still surveying the wider field rather than these two, we keep a broader roundup of the best Mac menu bar apps for tracking AI usage, and a tool-by-tool guide to tracking your Claude usage that covers the official Settings page as well. For a comparison against another paid menu bar tracker, see tokn.watch vs CUStats.