In March 2026, I was stopped by a usage limit in the middle of my work and finally opened Claude’s Settings page to look at the numbers. My extra usage had already reached 84%, and my seven-day usage was past halfway.
I thought that if the reminder stayed in view, I could at least plan the rest of my work earlier. So Claude and I worked together on a Chrome Extension called Claude Usage Nyan. An orange tabby sits in the toolbar, and clicking it opens the usage cards.
This article records that development experience as it happened. Later versions of the tool and the service interfaces may change, so the screens I saw then should not be treated as results that every account can obtain today.
What It Does
I put the reminder in three places: the toolbar badge, the popup that opens when it is clicked, and a floating status bar on the claude.ai page. The purpose was simple. I did not want to wait until work was blocked before remembering to check usage.
In the design at the time, the upper section showed account-side percentages such as the five-hour and seven-day windows, along with extra-usage fields. The lower section summarized the tokens and cost estimates that could be observed from conversations in that browser.
The meaning of cost needs to be clear. Multiplying estimated tokens by a model price list produces an API price-equivalent estimate. It is not an actual Pro or Max charge, and it is not a payment bill. A different price list, model identification, or estimation method can change the result.
Colors and icons can remind me to pay attention, but they cannot tell me how much work I can still finish. The numbers on screen still need to be read together with their source, update time, and error state.
Two Channels That Answer Different Questions
The first channel requests account data from an internal claude.ai usage endpoint. The public background.js shows that process and how fields such as five-hour, seven-day, and extra usage were parsed.
The data comes from the service side, but that does not make the endpoint a documented developer API with a stable public contract. If its fields, permissions, or response behavior change, the tool needs maintenance. A failed request also cannot be interpreted as zero usage.
The second channel observes conversation requests and streaming responses within the claude.ai page, then summarizes what that browser can capture. It cannot guarantee coverage of all activity on the account, and its estimate is not a complete record of the tokens used for actual billing.
The plan percentage above and the browser estimate below are signals with different scopes and different units. They can be observed side by side, but they cannot be directly subtracted or treated as errors measured against the same denominator.
Claude Code also needs to be discussed separately. Anthropic’s plan documentation says that Claude and Claude Code share plan limits when used through a subscription login, while API billing is a separate case. The browser interceptor cannot see individual CLI conversations, but that does not mean CLI use cannot affect the account-level plan percentage.
I wanted to keep both channels so I could see what each could provide, and what still remained unknown.
I Looked Around Before Building
Before writing the code, I looked at more than ten related tools. That sample included many native macOS apps, and at the time I did not find a combination that matched the interface and language needs I had in mind.
This was a limited search experience. It cannot establish that every tool was English-only, that other platforms had no choices, or that no one else had built an interface with two data sources.
For me, the product direction was still clear: present it as a Chrome Extension, put the two channels together, and start with Chinese. I later added English and Japanese so the interface could follow the browser language.
Those were choices made at the time. Supported environments and installation methods still need to be checked against the project’s current version information. A development journal cannot guarantee that every environment has been tested.
A Few Problems I Ran Into During Development (Skip if You’re Not Interested in the Technical Details)
The first problem was the response format. There was no public format documentation for the internal endpoint. When the popup first displayed the raw JSON, I gradually worked out which fields I needed. That helped with debugging, but it also reminded me that parsing a response once does not mean its format will never change.
Keeping debugging information can be useful, but a raw response may contain account or organization information. If I need to discuss a problem, I remove unnecessary fields instead of posting the entire response.
The second problem was the script execution environment. Chrome’s content script documentation distinguishes the page from an extension’s isolated world. I initially modified window.fetch inside the isolated environment and captured none of the page’s requests. I later moved the relevant script into the MAIN world and passed data back to the bridge layer through events.
That solved where the script ran. Entering the page’s execution environment also means the data boundary needs careful review. A working feature does not prove that isolation and security have been validated.
The third problem was the icon. It took three versions before I used a reference photo of an orange tabby and adjusted the ears and the M-shaped marking on its forehead until it was close to what I wanted. For this collaboration, a concrete reference helped more than continuing to describe the image in words.
The Limitations First
What matters more to me now is making the interface distinguish among account data, partial estimates, and read failures. A prominent percentage without its time window or basis can still mislead.
The login behavior and data flow also need precise language. The manifest from the same public version declares storage and alarms permissions, access to the claude.ai host, and page scripts. The background code uses credentials: 'include' so requests carry the existing authenticated state. That is different from claiming that the program directly extracts the value of a cookie.
Using local storage does not mean there are no outbound requests. The tool is designed to request data from claude.ai. Public source code makes inspection possible, but it cannot replace checking the particular version, its permissions, and the installed package.
For this revision, I did not sign back into an account to retest the tool, nor did I perform a complete security audit of a Web Store package or every dependency. What this article can explain is the development approach and part of the architecture visible in the public code. It cannot guarantee the current behavior of every account.
I always wanted a small tool that would remind me to look earlier. Clearly marking what remains unknown is closer to the tool’s original purpose than making every number look precise.
November 2026 revision: This revision preserves the March development experience while clarifying the data sources, estimation basis, authentication behavior, and the limits of what the public source code supports.


💬 Comments
Loading...