Checking Claude Code usage by project and day
What a token count can and cannot show was covered earlier. This piece is about how to look that count up by project and by day. Claude Code’s built-in usage or status command, and the account page, show how much the current session used, or how much quota the account has left in this period. Matching one repository’s local log from yesterday means looking somewhere else.
The log records input and output for each request
The detail left by each request is on your computer under ~/.claude/projects. It records the model used, the time, and how many tokens went in and came out. These files are logs that Claude Code itself writes on your computer. After you turn on AI usage metadata, Hina reads only the model, project, time, and tokens there. AI usage metadata is authorized separately and is off by default. It does not read what you asked.
In the log, the tokens for one request are split into input and output. Input is what gets sent to the model, and output is what the model writes back. Hina keeps the two separate when it totals them. When you check a day’s number, match input with input and output with output. When you have it read a whole draft in the blog repository and then revise it, input is often larger than output. Looking only at output can make the day look lightly used, even though that input belongs to the same day.
Named by repository root, with the day starting at 4 a.m.
After AI usage metadata is on, Hina names each project after its repository root. Editing back and forth in src and docs in the blog repository, or opening a worktree on the client site repository to fix one branch, still counts under its own repository’s name. An extra segment in the path does not create another project name. Sessions started in different folders of one repository also join that repository, split into rows only by tool and model, so add those rows together when you check that repository. Codex local logs are named by repository root as well, and what is read is again the model, project, time, and tokens.
A day is not cut at midnight. It runs from 4 a.m. to 4 a.m. the next day. If at 2 a.m. you are still in the client site repository and have Claude Code edit a page, that request counts toward the previous day. It is not split across two days just because midnight has passed. Other numbers Hina shows by day also start at 4 a.m. Viewed this way, a day may contain many requests. Added together, they are that day’s count, and the model used for each request is in that day’s record.
The input figure Hina shows already includes cache. When you compare it with the raw log, do not add the cache again. To check one day’s input for the blog repository, add up the requests under ~/.claude/projects from that day that belong to this repository, and compare.
The number is not a bill or a measure of work
What the tokens match is that model call in the local log. What they do not match is the bill. Hina does not estimate cost, and it does not multiply tokens by a unit price. Charges on the account page, and how much quota was used, still follow what Claude shows. With a different model, the same request to write an update note can use a quite different number of tokens. Comparing the blog repository’s numbers for two days in a row compares the scale of the calls, not which day changed more of the article.
The number is not the request count, much less the time you spent working on this project. One long session can use many output tokens while, in the middle of it, you read docs in the browser and edit other files in the editor. That stretch on the Trail stays recorded on the browser or the editor. High tokens do not turn it into longer Deep work. Few tokens on the blog repository for a day may mean most of the time went to reading code and organizing, in notes, the points to change. That work would not have entered the Claude Code log. A lot of output on another day does not mean the article is finished. Finished or not, you still have to open the repository and look at the changes left there.
To see where Claude Code was used yesterday, take the day cut at 4 a.m. and look at the blog repository and the client site repository separately. The total of all projects puts both repositories in the same cell. If, after you compare, it still does not match the quota on the account page, you can start by keeping the two sets of numbers apart. Quota belongs to the account and its billing period. Hina counts the tokens in this computer’s log, by repository and by day, with each day starting at 4 a.m. If the model changed midway through a session, or Claude Code also ran on another computer, the two sides will not show the same number.