Don't Work!

Your AI works. You don't.

日本語

How it works

Don't Work! shows how long your AI worked — your AI hours — from the logs Claude Code, Codex and Gemini CLI already keep on your computer. Here is everything it does, on one page.

1. What AI hours are

AI hours are the time your AI worked. They are counted from what your AI writes: 100,000 output tokens make 1 AI-hour.

A token is a piece of what the AI writes: about four letters of English, or one or two characters of Japanese. 100,000 of them is about a book's worth of writing.

Only what the AI writes counts. What it reads — your prompts, your files, its cache — does not. Chats in ChatGPT or claude.ai do not count either (they leave no record on your computer), and neither does OpenAI's dots: a dot works on its own cloud computer, so nothing lands on yours. Sessions running at the same time add up, so one day can hold more than 24 hours.

It is the time your AI worked, not time you saved.

2. What is read, and what is sent

On this site's page, for each reply of your AI, three things are read: the date, the tool (Claude Code, Codex or Gemini CLI) and the number of output tokens.

The tool your AI runs (the prompt way, in “How to connect” below) reads a little more, to count the kinds of work: the name of each tool the AI used, the extension of a file it edited (.ts or .md) and the first word of a Codex command. It sends only counts — per day and tool, how many times each of eight kinds of tool was used: writing code, writing text, reading, running things, researching, AI using AI, connected services and talking. To stop that, put "kinds": false in ~/.dontwork/config.json. To keep them out of the AI line on Feed, untick “Use my kinds of work in the AI line (shown on Feed)” on Settings. Neither takes back counts already sent: they stay on your records on Home.

Never kept or sent: your conversations, your code, the commands themselves, file names, paths, project names, model names, or how many AI accounts you have.

Nothing is sent until you press “Publish to this week's board”. Then one row a day goes out — the date, the tool, the output tokens and the hours — and nothing else.

The reading happens in your browser, on your own computer: the files are read on the page, and they are not sent.

3. How to connect

The easiest way: let your AI do it. Paste this prompt into the chat of Claude Code or Codex (the top of dontwork.co/connect has a button that copies it). Your AI does the rest in the terminal; you press “Allow” once, in your browser.

Connect this computer's AI records to Don't Work!. The instructions are at https://dontwork.co/agent — read them and follow them. Only ask me when the browser needs my approval.

However many computers you have, paste the same prompt on each of them and they add up on one account — see “I use several computers” under Common questions.

To do it yourself, counting the records in this browser:

  1. On the computer where Claude Code, Codex or Gemini CLI is installed, open dontwork.co/connect. It does not work on a phone.
  2. Press “Count my Claude Code records”, then “Choose the folder”, and choose the .claude folder in your home folder. It is hidden: on a Mac, press Cmd+Shift+H to go to your home folder and Cmd+Shift+. (the full stop) to see the hidden folders — or press Cmd+Shift+G and type ~/.claude; on Windows, type %USERPROFILE%\.claude into the address bar.
  3. If you have more Claude Code folders, such as .claude-acct2, press “Count my Claude Code records” again and add each one.
  4. For Codex, press “Count my Codex records” and choose .codex the same way.
  5. For Gemini CLI, press “Count my Gemini CLI records” and choose .gemini the same way (when it runs in the Mac's own sandbox, SANDBOX=sandbox-exec: ~/.cache/.gemini).
  6. Your AI hours for the last 7 days and in all, a 13-week band and a card appear. You can save the card as a PNG or post it on X.
  7. To go on this week's board, press “Publish to this week's board”. You sign in first, with X or GitHub.

4. How the board works

The board ranks one thing: AI hours, all your tools added together — by day, week, month, quarter, year, or all time. Money, raw token counts and levels play no part. Every period counts only the rows that arrived in time.

A week starts on Monday at 00:00 in your time zone: the board's "today" and "this week" are chosen by your device's clock, the same day as your home screen's. The counting under them (the entry ticket, the streak, the arrival deadline) goes by UTC days, so the same week is the same board for everyone; the API, the CLI and the AIs' posts use UTC weeks. There are three boards: World; Country, once you set your country in Settings; and Tool, which ranks by that tool's hours alone (World and Country add every tool together).

Only numbers that arrive in time count. A day counts if its numbers reached us within 48 hours of the day ending. A day has no cap — except what a tool older than 0.5.0 sent, which counts for 48 hours a day at most, because those readings could be wrong (syncing again with a newer tool lifts it for the days it confirms); when some time is not included, “Your place” on the leaderboard shows how much, which days and why. The first time you publish, every day you send counts, so you are on the board from the start of that week. (when the only rows you have are dated after UTC's day — a Japanese Monday before 9:00 — you join the board when UTC reaches that day)

On the leaderboard, under “Your place”, your week shows as “top n%”. See this week's board →

5. What GitHub is used for

Only to know it is you when you sign in.

Don't Work! asks GitHub for one permission: to read your email address. It does not read your repositories, your code or your commits.

Your GitHub username becomes your handle here. You can withdraw the permission at any time in GitHub's settings.

6. Hiding your name, deleting your records

In Settings, “Who can see my records” has three choices: “The world”, “Friends” and “Only me”.

To delete, press “Delete everything” under “Delete my records and my account” in Settings. Everything goes 24 hours after you ask; until then, you can call it off.

7. Common questions

Can the site see everything on my computer, or take my data?

No. It sees only the folder you choose yourself (.claude or .codex) and, if you add one, a Cursor usage CSV — nothing else on your computer. In that folder it counts only the date, the tool and how much the AI wrote (the output tokens); it does not read your conversations, your code or your file names. What leaves your computer is one row of numbers a day — the date, the tool, the output tokens and the hours — and, with a Cursor usage CSV, its day totals (the date, the output tokens, the requests, the edge mark) and the time zone, when you press “Publish to this week's board”; never the CSV file itself. To check for yourself: turn off Wi-Fi before you choose the folder, and it still counts; and before you publish, open “See what would be sent” to see the rows themselves. In Chrome and Edge the page also says “Permission on the folder: read only” under the numbers — that is what the browser itself answers.

Why does my browser ask whether to “upload” the files?

That is the browser's own wording when a page lets you choose a folder — not words Don't Work! wrote. This page sends no files: it reads them on the page. In Chrome and Edge the page now chooses the folder another way, and the browser asks whether to let this site view and copy the files instead. It also says the site can make its own copies: that, too, is the browser's wording — Don't Work! reads the files to count, and keeps or sends no copy. Safari and Firefox cannot choose a folder that way, so they still say “upload”.

When I sign in with GitHub or X, what can the site see?

On GitHub, it asks for one permission: to read your email address (user:email). It asks for no permission to read your repositories, your code or your organizations. On X, it asks to read your profile and your email address. X's own screen also lists viewing your posts: under X's rules, that permission comes with the one to read a profile. Don't Work! does not read your posts and does not keep them. It also asks to stay connected (offline.access), which lets it see nothing more. It never asks for permission to post, to follow or to read your DMs. What it keeps: your name, the link to your picture, your email and, with X, your X username. You can delete your records and your account on Settings at any time.

I don't know how to use the terminal. Is that all right?

Yes. Paste the prompt at the top of dontwork.co/connect into the chat of Claude Code or Codex. Your AI does the rest: it gets the tool, counts the records, sends them, and sets this computer to send every hour by itself. What you do is press “Allow” once, in your browser. What is sent is the numbers for each day — the date, the tool, the output tokens, the hours and how many times the AI used each kind of tool — and a random mark that tells this computer apart; nothing else (set "kinds": false in ~/.dontwork/config.json to send no kinds of work). The tool checks dontwork.co for a newer version once an hour, when it syncs, and replaces itself only after checking its signature; set "autoupdate": false in ~/.dontwork/config.json to stop that.

I use several computers. Can they all count on one account?

Yes. The steps are the same on each computer: paste the prompt at the top of dontwork.co/connect into that computer's Claude Code or Codex chat, sign in to the same account (X or GitHub) in its browser, and press “Allow”. Four computers, four times. Nothing has to be deleted first. A computer connected before this: to add it up, paste the same prompt there once more — it gets the newer tool and its own mark, and its numbers are not counted twice. The earlier days of a newly connected computer go into your record, but not onto the boards; what it works from the day it was connected is added. Your computers are listed on Connected computers, where you can remove one. How many you have, and each one's numbers, are shown to you alone. What the browser's own way (“Do it yourself”) sends carries no computer's mark, so to add up several computers, paste the prompt instead. You can give each computer a name on Connected computers (“Work MacBook”, say); the name is shown to you alone.

One Mac, several macOS users?

Once per user: paste the prompt into that user's Claude Code or Codex and sign in to the same account in that user's browser. Don't Work! counts each user as one computer and adds them up.

A week has only 168 hours. How can the board show more?

Because the h on the board is not a person's hours. It is AI hours: what the AI wrote, in output tokens, turned into time at 100,000 tokens = 1 hour. AIs running at the same time — several sessions, several computers — all add up, so one day can pass 24 hours and one week 168 (there is no cap on a day, except what a tool older than 0.5.0 sent: 48 hours a day at most). It is the time your AI worked, not time you saved. See 1. What AI hours are.

Do chats in ChatGPT, claude.ai or the Gemini app count? And OpenAI's dots? (Browser only — why 0 hours?)

No. Don't Work! reads only the logs Claude Code, Codex and Gemini CLI keep on your computer. A chat in the browser or the app leaves no log the reader reads, so there is nothing to count. The same goes for Codex users: only Codex's logs count — chats in ChatGPT, even on the same account, do not (and the same for claude.ai and Claude Code). OpenAI's dots are the same: a dot works on its own cloud computer, so nothing lands on your computer, and the Codex tasks a dot starts are cloud tasks that leave no record either. The Gemini app (gemini.google.com, the phone app) is outside what the reader reads too: it keeps no log the reader reads. Gemini CLI — Google's agent in the terminal — does, and its logs count from reader 0.8.0.

The numbers differ from the command-line reader.

They read different folders. The command-line reader (dontwork) reads every .claude*, .codex* and .gemini* folder in your home folder; this site reads only the folders you choose. Choose the others too.

A session that is running right now is missing.

Your browser will not read a file that is being written while it reads. Count the folder again later and it comes in.

How far back does it go?

Everything still on your computer is read — a machine that never deleted its logs had two and a half months on the day this was written. Claude Code has a setting that removes old transcripts (cleanupPeriodDays, 30 days by default): to keep them, add "cleanupPeriodDays": 3650 to ~/.claude/settings.json. Codex keeps its sessions. What you publish stays on dontwork.co, so the earlier you publish, the more piles up.

Does Cursor count?

Export your usage CSV by hand from Cursor's dashboard (cursor.com › Usage) and add it on /connect/: each day's output tokens go in. Only three columns are read — the date, the kind and the output tokens — and the file never leaves your computer. For Cursor's own Agent and Tab, we don't currently support a way to read them reliably from the logs on your computer.

What about GitHub Copilot?

Not yet: we don't currently count usage from GitHub Copilot. Claude Code, Codex and Gemini CLI are read today.

Who writes the “<name>'s AI” posts on the timeline, and why do they name other people?

Nobody. They are posts made by fixed rules from the person's numbers alone — today, this week, their place on the board, their streak — never from what they or their AI said or did. They name another person only when that person is public to the world and ranked on this week's board, and they say only what that person's board row and page already show.

What is the companion on the leaderboard, and how does it grow?

Anyone signed in can adopt a companion — a small pixel creature that lives on their Home. It grows with the AI hours in their logs (what the AI wrote), never with anything they press. The meal, the clean-up and the daily free spin cost nothing — the meal and the free spin even pay a game coin each on their day, and a day with an AI hour or more adds three coins to the pay on the following day or later; the coins (from those same AI hours) buy furniture and a bigger home. From October 26, 2026 the next stage of the home can also be taken at once on a coin loan: the price is the same as paying in full, the coins stay in the wallet, and it is repaid automatically from half of the hourly pay — in coins only. On the night of a day it was fed, it lays a mystery egg 20% of the time — coins or a gacha ticket inside. The spin gives wallpapers and transformation fruits, and a ticket is one more spin. Money plays no part: coins and tickets cannot be bought. The companion stands beside its person's row and on their public page; it plays no part in the ranking.

What is the AI work history?

A page of your own, reached from your public page, that fills in by itself from your logs: the days you have used AI, your streak, the tools, the hours — and, on your own screen only, the kinds of work. What you accomplished you add yourself, in your own words, one entry a day. It is a record, not a certificate: it counts what has reached dontwork.co, and it changes when records are corrected or a computer is removed.