Your token and value numbers will drop by roughly 2.5×. That is the fix, not a regression.
Clawdmeter was counting each assistant message once per content block. Claude Code writes one
message as several JSONL records — one for the thinking, one for the text, one per tool call —
and each repeats the same usage figures. Everything read from your transcripts was multiplied by
however many blocks a message happened to have.Your percentage bars and extra-usage spend were never affected — those come from Anthropic's
rate-limit headers and usage endpoint, not from transcripts. Only the token counts under the
bars, the per-session totals and the Stats page were wrong.
Downloads
| Platform | File | How to run it |
|---|---|---|
| Windows | Clawdmeter.exe |
Single self-contained file — just run it. |
| macOS | Clawdmeter.dmg |
Open, drag to Applications. Universal — one download for Apple Silicon and Intel. |
| macOS (zip) | Clawdmeter-macos.zip |
The same .app, if you'd rather not mount a disk image. |
| Linux | Clawdmeter-3.0.1-x86_64.AppImage |
chmod +x and run. One file, no install; delete it to uninstall. |
| Linux (tarball) | Clawdmeter-3.0.1-linux-x86_64.tar.gz |
Unpack, run ./install.sh — adds a menu entry. |
Every file has a .sha256 beside it. Full checksums:
879615136889f8436c48398005e8a4ea5a5d120eaaa79116b1b945e878761aaa Clawdmeter.exe
624de40071661358eb36dbdf20d578bd16eb95c18a574e6d11e2611ecacbaf7b Clawdmeter.dmg
cd9b779ac6e57b71ae112d025d149b471f6e4713f7344d10bd00e0760fbc7fbd Clawdmeter-macos.zip
ee5bdf383efd77688c66185b8805edb6a8429702618ac93cd182387594200e9d Clawdmeter-3.0.1-x86_64.AppImage
2ee59c25ab5efcdf18b789fc9370b5d0abb00c4ab4b8ae9c017eb87d571474b7 Clawdmeter-3.0.1-linux-x86_64.tar.gz
Verify with sha256sum -c <file>.sha256 on Linux, shasum -a 256 <file> on macOS, or Get-FileHash <file> on Windows.
None of these are code-signed yet, so each OS will question them once.
- Windows — SmartScreen → More info → Run anyway.
- macOS — Gatekeeper says "Apple could not verify Clawdmeter is free of malware." Try to open it, then go to System Settings → Privacy & Security → Security and click Open Anyway. Right-click → Open no longer works on macOS 15 (Sequoia) or later — Apple removed that bypass.
- Linux — no prompt.
Fixed
Token usage was counted per JSONL record instead of per message (#21). One assistant message is
written as several records — one per content block — each repeating the same message.id and the
same usage figures. Summing them counted every message once per block.
Measured across a real 18,654-message corpus: 73% of messages landed on more than one record, and
work tokens read 2.575× high. Reproduced on Windows, Linux and macOS transcripts — the reader
has no platform-specific paths, and the duplication is written by Claude Code itself, so every
install was affected identically, on every version before this one.
Two separate duplications are now resolved across the whole scan rather than per file: a message
split across several records, and a resumed session replaying those records into a second
transcript file. Records of one message are folded by taking the largest figure per bucket, because
output_tokens is a running total — early records carry a partial count — and a replayed copy can
carry an all-zero usage block.
What this changes on screen: the token counts beside the 5h/7d bars, the per-session totals on the
shelf, and everything the Stats page prices, including API value this month, lifetime value,
cache savings, value by model and project, and the turn count.
Finished the repo rename. The project moved to github.com/weltern/Clawdmeter on 2026-08-03,
but four files outside src/ still carried the old name — including the Windows version resource
you see in the exe's Properties dialog, and the AppStream metadata. The old URL only redirects, so
nothing was broken; it was just wrong.
Requirements
Windows 10/11 · macOS 13+ · Linux with glibc 2.35+ (Ubuntu 22.04+, Debian 12+, Fedora 36+)
Thanks
To @lizhuojunx86, who found the token-counting bug, measured it against a ~50-day corpus, and
filed #21 with a self-contained reproduction. The report was accurate, the diagnosis was right, and
the writeup made it straightforward to confirm independently and fix properly.
