Repository navigation
Releases
UsageAtlas 0.4.6 (unofficial fork)
UsageAtlas 0.4.6 (unofficial fork)
Quiet in the tray, and a provider directory you can actually use.
Settings → Providers
The catalog is now a real screen. Search all 305 tools the app knows, filter by trackable,
on, or installed on this PC, and see at a glance which ones have a reader and which are already
here. Real brand marks are used wherever the project ships one.
Nothing is enabled by a scan. Enable is the only switch, and a tool without a reader says so and
offers a local probe — it never claims to track what it cannot read.
Memory in the background
With power saving on, a window that stays in the tray for 30 seconds gives its renderer back to the
OS. The process, its heap and its compositor are released, and the window is rebuilt at the same
route and size when you open it again. The mini window is released the same way.
Between hiding and release the page paints at one frame per second and every CSS animation is parked.
Measured on the packaged build: 311 MB across four processes with the window hidden, against
535–579 MB with a window open. The renderer process is simply gone.
Background checks now follow what the app is doing:
| State | Usage check |
|---|---|
| Window visible | dashboard cadence (30–300 s) |
| Hidden, wall power | every 15 minutes |
| Hidden, on battery | every 30 minutes |
Discovery keeps looking
A remembered route is no longer trusted just because the file still exists: every file is re-probed,
the folder that held it is walked again for ledgers that appeared later, and sibling folders of the
same product are always checked. Results merge per path, and the discovery screen now reports
"confirmed by N files" next to the confidence.
Install
Download the Squirrel installer or the portable zip below. Windows x64, unsigned build; the app does
not require an account.
UsageAtlas 0.4.5 (unofficial fork)
UsageAtlas Unofficial 0.4.5
Four fixes, two more 0.5.0 slices.
Fixed
- The mini window no longer shows the page behind it. The card was 78 % opaque over a backdrop
blur, so Settings, another app or the desktop read through the glass. It is near-opaque now; only
the 8 px inset around it is transparent, so the corners stay round with no square edge. - The mini window can be resized. A transparent frameless window gets no resize border on
Windows, so the card draws eight handles and the main process follows the real cursor. 180 × 104
to 640 × 400, always inside the display, remembered between launches. - Cloud save sends the day it measured. A record whose model names are public is uploaded
unchanged — totals, hours and model rows. Only a private name such asm-094859is reported under
the plan's public model, because the server validates every model key against its catalog. A source
with no rule of its own follows the first saved mix, so Freebuff, Cline and Roo history can reach
the cloud once one plan exists. - Day details are plain JSON again. 0.4.1 deflated them; the next launch rewrites every deflated
row back to readable JSON and no new row is compressed. - The icon installs itself. The chosen mark is copied to one stable
app.icobeside the build
folders on every launch, andrefresh-install.ps1refreshes it and the Explorer icon cache after an
install, so a shortcut can no longer point into a folder an update deletes.
Added
- Installed tools, without reading a file. Settings → Find a provider scans each data folder once
and lists the catalog tools that actually have one; picking a name probes it in one click. - A hidden window does no work. The mini window stops its 30 s snapshot poll and the renderer
parks every animation while the window is hidden or minimized.
Checks
- 708 desktop tests pass, including new coverage for the resize geometry, the installed-tool scan,
and a save that keeps a day's full breakdown while relabelling a private model.
UsageAtlas 0.4.4 (unofficial fork)
UsageAtlas 0.4.4
A quieter mini window, discovery inside the app, and updates that tell the truth.
The mini window is one number
The floating window used to draw two quota dials that most local tools cannot fill, then spend a third
of its height on burn rate, session cost and a reset countdown. It now shows the only number it can
always prove: today's tokens across every enabled tool.
- Exact figures when there is room:
774,128,930. - Below 300 px wide it reflows to a short unit:
774.1M. - It can be resized down to 180 × 104 and up to 640 × 400.
- The provider logo and name stay in the header, with expand and hide buttons.
The square edge is gone. The native window is transparent, the Windows acrylic sheet that painted the
whole rectangle is removed, and the card is inset by 6 px inside the window. Its 22 px corners and its
shadow now follow the card, not the window.
Discovery has a screen
Settings → General → Find a provider accepts any tool name and probes the files that name owns on
this computer. The suggestion list carries the 63 seeded provider names. A scan reports:
- the route it found, or the route it remembered from an earlier scan;
- a confidence score;
- the number of folders and directory entries scanned;
- every matching file, with its store (
json · usage.total_tokens) and the token fields it saw.
No network request is made. This is the first slice of the 0.5.0 “any tool, 30 s” goal; the 300+
verified provider catalog and the 98 % installed-tool detection target remain open.
Updates: restart badge, rollback kept
- When Squirrel has unpacked a newer
app-<version>while an older window kept running, the top bar
now says Restart to finish update, not “New update available”. - The update state re-reads the install root on every read, so a stale “latest” cannot survive a newer
folder appearing on disk. refresh-install.ps1keeps the newest build, its immediate predecessor and the folder the running
window came from. The one-click Roll back to therefore still has a target after an
update instead of finding an empty disk.
This is the first slice of the 0.5.0 “trusted by construction” goal; signed packages, an encrypted
vault and an SBOM remain open.
Build
Windows x64, per-user, no administrator prompt. The build is not Authenticode-signed, so SmartScreen
asks for More info → Run anyway. SHA256SUMS and the Squirrel RELEASES/-full.nupkg feed sit
next to the installer.
UsageAtlas 0.4.3 (unofficial fork)
UsageAtlas 0.4.3
A cost estimate that looks like a bill, and a capacity card that works for every tool.
$0.54 per million tokens
The old estimate read $5.53 for 774 M tokens because Freebuff stores private model names such as
m-096e75164d: the catalog price for the resolved public model is far below what the subscription
actually costs. Settings → Data now has a blended cost estimate, on by default at $0.54 per
million tokens. That is the rate at the top of the public leaderboard — 42.29 B tokens for roughly
$23 k — so a local total and a cloud profile finally tell the same story.
The blended rate wins over catalog prices and custom overrides while it is on. Turn it off to go back
to per-model pricing. Applying the rate rewrites every stored day and observation, not only new
requests, so existing history moves too. On a measured 762 M-token profile the stored estimate went
from $82.88 to $411.46, with every token count unchanged; the 774 M total displayed on screen
moves from about $5.53 to about $418.
For 169 M tokens it lands near $91 instead of $70, at the leaderboard's rounded rate.
Repricing is idempotent: the second launch finds every row already at the rate and writes nothing.
Active tool capacity
Today's capacity card used to show an empty dial when a provider reported tokens but no quota window.
Freebuff and Claude do exactly that locally. The card now lists each enabled tool with:
- its logo and name;
- today's token count and estimated cost;
- its plan when the provider reports one, or No quota reported.
Tools that do report limits keep their existing gauges, and any tool without a limit appears in an
Also active row below them.
Repository and update hygiene
- The updater now hands the single-instance lock to the newest installed build instead of refocusing
the previous one, andrefresh-install.ps1removes olderapp-*folders and refreshes Explorer's
icon cache on every launch. - Roughly 1.2 GB of superseded installers, build output, diagnostic scripts and logs were removed.
Only the current release artifacts remain. - The discovery catalog is documented at its real size: 63 seeded provider names. The 300+
verified-provider target now belongs to the 0.5.0 contract, which also grew to seven acceptance-
tested outcomes: every tool in 30 seconds, bill-exact cost, a year on three machines, invisible
load, a trusted build, twelve languages and one million requests that stay instant.
Build
Windows x64, per-user, no administrator prompt. The build is not Authenticode-signed, so SmartScreen
asks for More info → Run anyway. SHA256SUMS and the Squirrel RELEASES/-full.nupkg feed sit
next to the installer.
UsageAtlas 0.4.1 (unofficial fork)
UsageAtlas 0.4.1
Attribution returns, and the local file is half its size. The hosted service still does not
accept every source, so a source it does not model can be reported as official providers again —
while this computer keeps the provider that actually produced the request.
Attribution, while the provider PR is unmerged
The Attribution page is back. A rule picks one or more official providers with a share each
(30 % Codex, 20 % ChatGPT, 50 % Claude, say), and every request draws from that wheel using a hash
of its own identity: the shares hold over time, the same provider can come up several times in a
row, and a save that is retried keeps the label it already had. Local history is never rewritten —
only the uploaded copy is. A source with no rule stays on this computer under its real name and is
listed in Settings → Cloud save.
The chosen target is now stored on the row it describes, so it disappears with that row and cannot
grow on its own.
Storage: 1,380,352 → 704,512 bytes
Measured on a copy of a real profile (20 day records, 606 observation rows, 723 labels):
| Before | After | |
|---|---|---|
| Database | 1,380,352 B | 704,512 B (49 % smaller) |
| Observation rows | 606 | 278 |
| Label settings rows | 723 | 0 |
Day records, tokens, costs, hourly buckets and per-model splits are identical. What changed:
- attribution labels are a column on the fact or day they describe instead of a settings row keyed
by its id; - two indexes no query filtered on (
model,project) are dropped; - a saved request keeps its payload for a week instead of a month — its fingerprint stays, so a
rescan still uploads nothing twice; - the one-time
VACUUMcould never run because the background writer opened a second connection at
startup; the writer now opens on its first batch, so the rewrite actually happens.
Every setting from the reference list
Settings gained the controls that were constants before: start minimised, global hotkey, background
price sync, power saving, window opacity and blur, density, tray metric, per-provider colours,
custom log roots (including mounted or WSL paths), ignored projects, stale-session timeout, model
price overrides, currency and fixed rate, tax, prompt caching, burn-rate and quota-reset alerts,
do-not-disturb and notification sounds, project redaction, retention, settings backup, encrypted
exports and a scheduled export. The compact always-on-top widget has its own page.
Build
Windows x64, per-user, no administrator prompt. The build is not Authenticode-signed, so SmartScreen
asks for More info → Run anyway.
This build also stops shipping Electron's 50 unused locale catalogs (47 MB installed); only the five
languages the interface offers are packaged.
SHA256SUMS and the Squirrel RELEASES/-full.nupkg feed sit next to the installer.
UsageAtlas 0.4.0 (unofficial fork)
UsageAtlas 0.4.0
The honest release. Cloud saves now carry the provider that actually produced the usage. The
weighted attribution wheel and its relabelling rules are gone, including rules still stored by
0.3.x; a source the hosted service does not model stays on your computer instead of arriving under
another provider's name.
Storage: 4.61 MB → 0.856 MB
Measured on a copy of a real profile with 259 requests:
| Before | After | |
|---|---|---|
| Database | 4.61 MB | 0.856 MB (18.6 %) |
| Observation rows | 2,856 | 285 |
The day record per provider is unchanged and remains the source of every screen. What left:
collectionandcapacitysnapshots keep only their newest copy instead of one per scan/check;- cloud-saved request payloads older than 30 days keep only a content fingerprint, so a rescan still
uploads nothing twice; - the write-only
usage_record_historyarchive is dropped; - one
VACUUMreturns the freed pages.
Totals, hourly buckets, per-model splits and estimated costs are identical. Restore no longer copies
the cloud observation ledger back into SQLite.
Also in this release
- 250-provider discovery catalog — offline folder spellings and content probes, unknown name
returns empty in under two seconds. - Notification inbox — the last 50 alerts with arrival time, kept across restarts in the
Notifications tab, even when Windows hides the toast. - Taskbar and toast identity — the Start Menu shortcut behind the AppUserModelID is written on
every launch with the chosen ICO, before the first window is created. Restart once after updating. - One-click rollback — Settings relaunches the newest older Squirrel build; local history is a
separate file and is untouched. - README rebuilt from upstream with the 0.4.0 measurements and the 0.4.0 → 0.5.0 timeline.
Install
Windows x64, per-user, no administrator prompt. The build is not Authenticode-signed, so SmartScreen
asks for More info → Run anyway. The portable ZIP needs no installer.
SHA256SUMS and the Squirrel RELEASES/-full.nupkg feed sit next to the installer.
0.5.0
Seven acceptance-tested goals are listed in the README: deeper appearance, complete translation,
cheaper idle, discovery in the app, a compact cloud format if the hosted service accepts it, a
guarded migration and trusted updates. No date is promised.
UsageAtlas 0.3.9 (unofficial fork)
UsageAtlas 0.3.9 — Windows x64.
This is an unofficial fork, not affiliated with, endorsed by, or supported by UsageAtlas. It is
based on SRX9/UsageAtlas and adds private provider readers and
an attribution layer that lets their usage be reported as official providers.
0.3.9 is the last release before 0.4.0. Code for the 0.4.0 goals lands on main as it is written,
but no 0.4.0 build ships until every acceptance test in the README is met.
Fixed
- Cloud save no longer looks stuck on sources the server refuses. The page counted every pending
observation, including Cline and Roo, which the hosted service does not model and no attribution
rule renames. They can never upload, so the counter stayed above zero and Save to cloud appeared
to do nothing. It now counts only what a save can send, reports the held-back sources with their own
total, and says Everything the server accepts is saved to cloud when nothing uploadable is left.
The hourly timer also ignores held-back rows instead of waking for them. - One copy of the app runs at a time. Squirrel leaves the previous build running after an update;
several versions shared this profile, competed forhistory.sqlite, and the taskbar group kept the
first build's icon. A single-instance lock focuses the open window instead of starting another copy. - Windows gets the ICO, not the PNG. The window is created with the selected mark's ICO, so the
taskbar button draws the app's artwork instead of the Electron atom.
Changed
- The local archive is bounded.
usage_record_historywas written on every collection pass and
read by nothing, growing to 4.2 MB of an 8 MB database in two days. It keeps the newest 20 versions
per record, on write and on open, then vacuums the freed pages. On a copy of a real profile the file
went from 7.91 MiB to 3.99 MiB. - README gained the 0.4.0 storage budget. It explains the one-document-per-provider-and-day format,
its additive merge rules, idempotence and trade-offs, with a fake-data comparison: 31,050 current
observations (30.1 MB) become 150 day documents (0.29 MB), 1.0 % of the payload and 207× fewer rows.
A provider table lists the tools the discovery work targets and where their token counts live.
Verified
- 678 desktop tests pass, including new coverage for the held-back counter, the timer staying idle for
unmapped sources, and the archive bound on open and on write; typecheck and lint are clean. - A smoke start of the packaged build initializes the window, the preload API and the renderer.
Notes
- The build is unsigned: Windows SmartScreen may warn on first run.
- Requires Windows 10 or newer, x64. Settings, history and route caches are kept on upgrade.
- Usage the server cannot store without an attribution rule stays on this computer; add a rule on the
Attribution page to report it as an official provider.
UsageAtlas 0.3.8 (unofficial fork)
UsageAtlas 0.3.8 — Windows x64.
This is an unofficial fork, not affiliated with, endorsed by, or supported by UsageAtlas. It is
based on SRX9/UsageAtlas and adds private provider readers and
an attribution layer that lets their usage be reported as official providers.
Fixed
- A cloud save uploads every pending change in one run. The hosted service stores at most 40
observations per request; the client sent its store's 250-fact page and expected all of them back,
so it failed withIncomplete statistics acknowledgementafter the first 40 and the counter came
down 40 at a time. Each request is now capped at 40 and the loop continues until nothing is left. - The Settings tab bar lines up with its cards. It was inline-level, so the auto margins that
centre the page blocks did nothing and the segmented control started at the page gutter. - The chosen app mark reaches the Windows taskbar and Start Menu shortcut. The window now also
publishes its icon throughsetAppDetails, rewrites the shortcut icon, and asks Explorer to refresh
its icon cache. The tray and title bar were already following the picker. - Alerts no longer disappear with a hidden toast. Every system notification is mirrored as an
in-app banner with the same text, a link to the right page and a dismiss button, so Focus assist or
a per-app Windows switch cannot swallow it.
Added
- A visual timeline in the README.
docs/timeline.svgshows every shipped 0.3 release and the
seven 0.4.0 goals in the app's own dark glass style, each goal with its acceptance test.
Verified
- 675 desktop tests pass, including a save of 50 observations through a server that acknowledges
only the first 40 per request, plus typecheck and lint. - The packaged renderer carries version 0.3.8, the centered tab bar, the alert banner and the new
taskbar icon code.
Notes
- The build is unsigned: Windows SmartScreen may warn on first run.
- Requires Windows 10 or newer, x64. Settings, history and route caches are kept on upgrade.
UsageAtlas 0.3.7 (unofficial fork)
UsageAtlas 0.3.7 — Windows x64.
This is an unofficial fork, not affiliated with, endorsed by, or supported by UsageAtlas. It is
based on SRX9/UsageAtlas and adds private provider readers and
an attribution layer that lets their usage be reported as official providers.
Added
- Four more appearance settings. Settings → Appearance also offers the number style, an accent
tint for buttons and focus rings (theme, ink, cobalt, violet, rose, forest, amber), glass or solid
card surfaces, and a switch that turns interface motion off while progress spinners keep turning.
Every choice applies immediately and is saved locally. - Full figures by default. Totals, cards and tables print every digit —
1,284,930instead of
1.3M— with Short units one click away. Chart axes stay compact in both modes so labels never
collide, and currency follows the same choice. - A restart prompt when the installed build is newer. An update applied while an older window is
still open now reports Version X is installed, and updating relaunches through the Squirrel stub
at the install root so the window ends up on the build actually on disk.
Changed
- The Appearance preview reads its figure from the same formatter as the dashboard, so the sample
follows the chosen digit style. - README gained a checkable 0.4.0 plan: provider discovery for unmapped tools, compact cloud saves,
deeper appearance controls, per-model attribution with a preview, trusted updates with rollback,
complete translation, and cheaper idle.
Fixed
- A stale window could claim You're on the latest version while a newer release was already
unpacked next to it. Squirrel compares against the package on disk, so the older build now reports
the waiting version instead of the newest one.
Verified
- 673 desktop tests pass, plus typecheck and lint.
- The packaged renderer bundle carries version 0.3.7, the new catalog keys and the accent, surface,
motion and number settings.
Notes
- The build is unsigned: Windows SmartScreen may warn on first run.
- Requires Windows 10 or newer, x64. Settings, history and route caches are kept on upgrade.
UsageAtlas 0.3.6 (unofficial fork)
UsageAtlas 0.3.6 — Windows x64.
This is an unofficial fork, not affiliated with, endorsed by, or supported by UsageAtlas. It is
based on SRX9/UsageAtlas and adds private provider readers and
an attribution layer that lets their usage be reported as official providers.
Added
- The rail highlight slides between pages. The active-page pill is a single element moved between
rail buttons, so going from Today to Profile — or from Profile to Usage alerts — glides down the
rail instead of jumping. It is measured from the real buttons, so it stays aligned when the window
or the type scale changes, and it is placed without animation under reduced motion.
Changed
- The appearance pickers are translated too. The font, app-icon and ink descriptions in
Settings → Appearance no longer stay in English inside a French, Spanish, Chinese or Hindi page,
and neither does the Custom background label. - The chosen app mark reaches the Windows taskbar. Each mark ships a multi-size ICO beside its
PNG and the window icon is set from it, so the taskbar and Alt-Tab artwork follow the picker
instead of only the title bar and the tray. Windows keeps the icon of the installed shortcut
itself; only the running window can change it. - French uses the domain word tokens rather than translating it to jetons.
Verified
- 667 desktop tests pass, plus typecheck and lint.
- The packaged renderer bundle carries all five catalogs and version 0.3.6.
Notes
- The build is unsigned: Windows SmartScreen may warn on first run.
- Requires Windows 10 or newer, x64. Settings, history and route caches are kept on upgrade.