The Robot Read My Claude Quota With 235 Billion Parameters. I Fired It and Hired a JSON File.

In February I introduced you to the robot: a PowerShell script that mashed Print Screen every sixty seconds, and a 235-billion-parameter vision model that read my Claude quota percentages off the screenshots like a very expensive grandmother squinting at a thermometer.

It worked. That was never the problem. The problem was that it worked, and therefore I had no reason to stop, and somewhere in Frankfurt a GPU was burning watts to tell me the number 68.

Æ brukte to hundre og trettifem milliarder parametre på å lese fire tall.

The robot is retired now. This is the story of its replacement, which is a JSON file with a timestamp.

The confession

There was an endpoint all along.

Claude Code logs you in with OAuth. The token it stores can askapi.anthropic.com/api/oauth/usage for your quota, and the answer comes back as clean, well-formed JSON — session percentage, weekly percentage, reset times, the lot. No screenshots. No OCR. No vision model contemplating my browser tabs.

I did not discover this. Smarter people did —claude-quota-tracker,ccusage, andccstatusline-usage all got there first, and credit lives in the README where it belongs. My contribution was noticing that every one of these tools assumed you lived a tidy little life inside one operating system.

I do not. I run Claude Code natively on Windows 11 and inside WSL2, sometimes in the same hour, because being from Northern Norway apparently means never having to say "this environment is good enough."

The architecture (now with actual architecture)

claude-code-quota is two small libraries — one PowerShell, one bash — that your status line script calls on every refresh:

1. Check the age of a local cache file (~/.claude/quota-data.json)
2. Fresh enough? Return immediately. No network call, no waiting.
3. Stale? Fire a background refresh — detached, non-blocking — and return the previous frame's data right now. The next frame gets fresh numbers.
4. First run ever? Block briefly, once, so the status line isn't blank.

The TTL is 60 seconds while you're actively working and 5 minutes while you're not, which the library figures out from how recently your session transcript was written. No daemon. No scheduled task. No process squatting in the background waiting for you to need it. The status line refresh is the scheduler, and it never pays more than the cost of reading one small file.

The result, permanently in view:

Sonnet 4.6 | main | ctx:42% | 5h:68% ~1h12m | 7d:31% ~4d2h

Windows was the hard part, obviously

Here's the trap: Claude Code on Windows runs your status line command through Git Bash. Git Bash does not ship jq. Every bash-based quota tool therefore opens with "step 1: install jq into Git Bash," which is the step where half your Windows users quietly close the tab.

So the Windows side is pure PowerShell — Invoke-RestMethod andConvertFrom-Json, both built in, zero installs. Cold start is roughly 100–200 ms on PowerShell 7 and 300–500 ms on 5.1 (measured casually on my machine; don't build a datacenter around it).

And because both libraries write the same cache file in the same format, Windows and WSL2 share one quota cache. Native session fetches it, WSL2 session reads it for free. As far as I can tell nobody else in this particular sandbox handles the dual-environment case, which is either a market gap or a sign that everyone else has healthier work habits.

Today's episode: the warning sign that rendered as doubt

This very evening, my status line said:

5h:29%? ~50m

A question mark. Was the data stale, or was the stale-marker glyph broken?

Both. Obviously both.

The data was stale — fetches had been failing for a while and the library was honestly flagging it, exactly as designed. But the flag is⚠, and what I got was ?, because when PowerShell's stdout is captured through a pipe — which is precisely how Claude Code runs your status line — it encodes output with the OEM code page. A character set standardized when the fax machine was aspirational technology. It has no ⚠, so .NET shrugged and substituted a question mark, turning a warning into an existential one.

The fix is one line at the top of the status line script:

[Console]::OutputEncoding = [System.Text.Encoding]::UTF8

That warning glyph had been silently eaten on every Windows install since day one, and nobody noticed, because noticing required your quota fetch to fail and your typography standards to be unreasonably high at the same time. Mine were. It's fixed in v1.0.0.

v1.0.0, a plugin, and other signs of adulthood

The repo now has an actual tagged release, and it's installable as a Claude Code plugin — the repo is its own marketplace:

/plugin marketplace add aweussom/claude-code-quota
/plugin install claude-code-quota@aweussom

That gives you a /quota skill immediately: ask Claude Code for your quota and it reads the cache and tells you, including how stale the data is and which version wrote it. Plugins can't touch your statusLinesetting (that's yours, in settings.json), so for the permanent display you run the installer once:

# Windows
pwsh -ExecutionPolicy Bypass -File .\install.ps1


# Linux / WSL2
bash install.sh

Or ask Claude to "set up my quota statusline" after installing the plugin, and the skill walks it through the whole thing. We live in an age where the installation instructions can install themselves. I have decided not to think about this too hard.

The honest part

• The endpoint is undocumented. The library pins a beta header (oauth-2025-04-20), and Anthropic can change any of it on any given Tuesday. If that happens, the status line degrades to showing stale data with a warning glyph — a visible one now — and I bump the header. That's the deal. That's also why there's finally a version number in the cache file.
• Claude Code has /usage built in, and it's fine. If you check your quota twice a week, you don't need me. This tool is for the people who want the number continuously in view without a daemon, across Windows and WSL2 at once.
• If what you actually want is cost analytics — tokens, models, spend over time — ccusage is genuinely excellent and you should use it. Different job.
• Not affiliated with Anthropic. They just make the thing I keep running out of.

Things I learned

1. Before building a robot to read a number off a screen, spend one hour checking whether the number is available as a number.
2. Fresh-with-a-timestamp beats fresh-on-demand for anything a status line does. Return the old value, refresh in the background, and nobody ever waits.
3. If your output can contain anything invented after 1987, set your encoding explicitly. The OEM code page outlives us all.

The repo

aweussom/claude-code-quota— MIT licensed, no dependencies on Windows, jq and curl on Linux. Star it if you too have been personally victimized by Claude rate limits, or if you just enjoy watching a warning glyph render correctly.

The robot's repo stays up as a monument. We do not delete our origin stories; we link to them sheepishly.

---

The author is a Systems Specialist who does not work in software development. His Claude quota is now read by a cache file at a total cost of zero parameters, and he has redirected the 235 billion he saved toward asking Claude why the quota is always at 68%. He is not expecting a satisfying answer.