Your Claude 5-hour and weekly usage limits, live on the 128×128 screen of an Epomaker × Aula F75 Max.
curl -fsSL https://raw.githubusercontent.com/schiz0x00/keyboard-screen/main/install.sh | shGrabs the latest release binary, verifies it against the release checksums (or
builds from source if there's no release for your platform), installs it to
~/.local/bin, and adds the udev rule so the screen is writable without root.
Replug the keyboard afterwards.
To override anything, put the variable after the pipe — before it, the shell
would set it on curl instead:
curl -fsSL .../install.sh | KEYBOARD_SCREEN_BIN_DIR=~/bin KEYBOARD_SCREEN_SKIP_UDEV=1 shkeyboard-screen # refresh once and push to the screen
keyboard-screen --watch # keep it updated, default 60s interval
keyboard-screen --watch 300 # ...or a slower one (60s minimum)
keyboard-screen --dry-run -o /tmp/f.ppm # render without a keyboard attached
keyboard-screen --no-refresh # draw from cache; spend nothing| Option | What it does |
|---|---|
-w, --watch [SECS] |
Keep refreshing. The interval is optional and defaults to 60s; anything below that is raised to 60s, because each refresh spends quota. |
-s, --slot N |
Display slot to write, 1–255 (default 1). |
-n, --no-refresh |
Never ping the API; draw from the cache or Claude Code's file. |
-o, --preview FILE |
Also write the frame as a binary PPM for inspection. |
--dry-run |
Render but don't touch the keyboard. |
--test-pattern |
Push an unmistakable test frame instead of the dashboard. |
-v, --verbose |
Log what it is doing, including why a refresh fired. |
-h, --help, -V, --version |
Help and version. |
(--wireless, --probe-wireless and --probe-write exist only for the
reverse-engineering write-up; the screen is wired-only. See
docs/wireless-re.md.)
In --watch mode the API is pinged on the interval, but the screen does not
wait for it: between pings the loop reads Claude Code's own
~/.claude/usage-exact.json every two seconds - a free local read - and pushes
a fresh frame within seconds of the numbers climbing, at most once every 15
seconds. When nothing is being spent, it stays on the plain interval.
Run it continuously:
install -Dm644 systemd/keyboard-screen.service ~/.config/systemd/user/
systemctl --user enable --now keyboard-screenFrom the anthropic-ratelimit-unified-* headers of a ~9-token Haiku ping,
authenticated with the OAuth token Claude Code already stores in
~/.claude/.credentials.json.
This spends subscription quota, never API credits. The only credential it
will send is a Claude Code OAuth token (sk-ant-oat…); it never reads
ANTHROPIC_API_KEY, never sets an x-api-key header, and refuses anything it
doesn't recognise rather than guessing which account pays.
Two things worth knowing: the headers are account-wide, so the screen also reflects usage from claude.ai or another machine — which is the point; and a ping opens a 5-hour window if none is running, so watching around the clock keeps one open.
Claude Code's own usage-exact.json is a last-resort fallback only — it
refreshes when Claude Code feels like it, not when you ask. The footer always
says which source the numbers came from and how old they are (LIVE,
CACHED 4M, CC FILE 2H30M), so a frozen screen can't pass for a live one.
The screen is wired-only. Over 2.4GHz or Bluetooth the tool exits cleanly and the screen keeps its last frame, correctly labelled as stale — so you can leave the service enabled full time.
docs/protocol.md— the screen protocoldocs/wireless-re.md— why wireless can't work, with the decompiled proofCONTRIBUTING.md— how to build, test and release
cargo build --releaseDependencies are libc (raw ioctl) and ureq + rustls for the HTTPS ping.
The JSON parser, 5×7 font, drawing primitives and HID transport are all in-tree.
MIT — see LICENSE.
The screen protocol was reverse-engineered by the Aula-F75-Max-Driver and Aula-F75-Max-OSX projects; this implementation confirms it independently against the vendor software.