1. How much quota is left, and when does it reset? Every provider shows usage in a different place.

2. Is each CLI up to date? They update often, and it is hard to keep track.

3. Which model gives the best value? The highest-scoring model is not always the one that fits your budget.

I built three tools, one for each question. All are open source under the MIT license:

Both skills cover Claude Code, Codex (ChatGPT), Grok Build, Muse Code and Antigravity (Gemini). You don't need all five; a CLI that isn't installed shows not installed.

1. ai-usage: one report for five providers' quotas

After installing, type this in Claude Code, Codex or Grok Build:

/ai-usage

A sample report. Quotas, versions and rankings in the screenshots are examples; actual values depend on when you run it.

The report shows:

• Usage and reset times: Claude, ChatGPT and Muse have a 5-hour and a weekly window; Grok has only a weekly window; Gemini has two weekly pools.

• now: the model and effort (reasoning level) each CLI used in its most recent local session.

• Qualified: for each provider, the model and effort with a CursorBench score of at least 50% and the best CP (score ÷ cost per task); if nothing reaches 50%, the highest score is shown and marked *. If CursorBench doesn't list the model, Artificial Analysis (AA) data is used instead and marked (AA). These are the same two sources AI Model Compare uses (section 3), but the web page defaults to AA.

• ⚠: shown only when a weekly window is over 80% used.

• →: which provider to use now, with a one-line reason.

CP = score ÷ cost per task, and is only compared within one benchmark. Cost is based on API prices, not your subscription fee, and does not show how much subscription quota is used.

Design notes

Quotas come straight from the CLIs you are already signed in to, without driving a browser or reading passwords or cookies; benchmark data is fetched separately from public pages. Running it in Codex still shows your Claude quota.

• Missing data is never treated as zero. A failed call shows an error. Grok omits its usage percentage below 1%, so the report shows <1%.

• Checking Muse uses a little quota. It has to send a very short request (Reply with the single word: ok) to get usage. This takes about 15 seconds and leaves one test session.

• Report language is configurable. Ask for a language explicitly, for example "answer in Traditional Chinese"; otherwise it uses your system's preferred language when available, or English. The script's raw output stays in English.

2. ai-cli-version: are your five CLIs up to date?

/ai-cli-version

In this example only Claude Code has a newer version. Each row shows the installed version, install time (approximate), latest version, release time and status.

The check only looks things up; it never installs or updates a CLI on its own. When updates are available, the agent asks: y updates every outdated CLI, Esc skips them all. Running the Python script directly only checks.

Version data comes from the npm registry and each vendor's update source. Grok uses --check, and Muse and Antigravity have their auto-update on launch turned off. The skill doesn't need npm or Node.js to run.

Install times come from file timestamps and may be off. Grok's release time comes from a third-party mirror and is marked ≤; Muse and Antigravity publish no release dates, so they show ?. If you installed a CLI with Homebrew or npm, update it with that package manager.

3. AI Model Compare: model score vs. cost per task

Open the web page; nothing to install. Data refreshes automatically every 6 hours (the footer shows when), and the previous data is kept if a refresh fails.

A 70-second walkthrough:

https://www.youtube.com/watch?v=CB1NLN4oAXw

Chinese version: watch on YouTube

Inspired by the CursorBench leaderboard, it plots score against cost as a scatter chart. It uses Artificial Analysis (AA) by default, and you can switch to CursorBench:

• X axis: API cost per task; cheaper is to the right.

• Y axis: score by default; higher is better.

• Each point is one model at one effort level, and a model's effort levels are joined into a line.

The page opens on the Artificial Analysis (AA) Intelligence Index. Use "Source" to switch, or open CursorBench mode directly.

The same chart on CursorBench. CursorBench doesn't list GPT-6.1 Sol yet, so the chart shows GPT-5.6 Sol for Codex. With score on the Y axis, the upper right means higher score at lower cost.

The two sources use different test sets, so their scores and CP can't be compared with each other; the AA index is not a percentage either.

How to use it

Hover over a point and the other lines fade, with crosshairs marking its values. Here Opus 5.5 High is selected; the readout below shows AA's Intelligence, Cost, Speed, Verbosity and Latency. On CursorBench it shows Score, Cost, Tokens and Steps.

Click the ▾ next to a legend chip to tick models and versions one by one, or click a line's name to tick the whole line. On first load each provider shows the newest version of one model; older versions are drawn paler.

Click a legend chip to show or hide a whole CLI; here only Codex is visible. Turning a CLI back on keeps the models you ticked. The number on a chip is its point count, shown as "shown/total" when some models are unticked. These settings are saved in your browser, so a reload or your next visit keeps them; each data source keeps its own ticks.

The Y axis can also show AA's Speed, Verbosity and Latency, or CursorBench's Tokens and Steps. Higher speed is better but lower latency is better, so the upper right isn't always best after switching.

Installation

You need Python 3.9 or newer; no pip install. Before checking quotas, install and sign in to at least one supported CLI, and the agent must be able to launch local programs and reach the network. Checking versions doesn't need a sign-in.

The skills CLI currently requires Node.js 22.20 or newer. To install both skills into Claude Code, Codex and Grok Build at once:

npx skills add Jeinn-co/agent-skills -g -a claude-code codex grok -y

To install just one:

npx skills add Jeinn-co/agent-skills --skill ai-usage -g -y
npx skills add Jeinn-co/agent-skills --skill ai-cli-version -g -y

-g installs globally; -a limits the install to the agents you use. On Windows you can add --copy. After installing, open a new session and type /ai-usage or /ai-cli-version.

To update:

npx skills update ai-usage -g -y
npx skills update ai-cli-version -g -y

Open a new session after updating. If you'd rather not install Node.js, copy the skill folder into ~/.claude/skills/, ~/.codex/skills/ or ~/.grok/skills/, and copy it again to update.

Platforms

Tested on macOS and Traditional Chinese Windows 11, though not every CLI and platform combination; Linux hasn't been tested yet. If you run into a problem, please open an issue with your OS, CLI version and output.

Links

• Skills source code

• AI Model Compare source code

• AI Model Compare web page

If these tools help you, a star on GitHub is appreciated; for problems or support for another CLI, please open an issue.