An Antigravity usage tracker earns its place by answering two questions well: which model consumed the week, and which session was the expensive one. Anything beyond that, such as remaining plan capacity, is a different problem with a different owner. The UsageAtlas integration guide draws that line explicitly. Ranking the available approaches means ranking how honestly they draw it.
The ranked options
| Rank | Approach | Reads | Honest about limits |
|---|---|---|---|
| 1 | UsageAtlas with Antigravity CLI enabled | Local .db conversation history |
Yes, states it has no quota data |
| 2 | The CLI's own history or status output | Session-level information | Depends on what it prints |
| 3 | A spreadsheet you fill in weekly | Whatever you type | It is your own estimate |
| 4 | A general log parser | Files you point it at | Usually silent about scope |
UsageAtlas takes the top spot for two reasons: it reads the format the CLI actually writes, and its documentation says what it cannot see.
What the reader parses
The CLI writes conversation data to a directory under the home folder, and UsageAtlas reads supported .db files from it. The provider documentation lists alternate locations and an environment override.
Two things do not work, and both are documented rather than discovered.
- IDE history. The Antigravity IDE stores
.pbfiles, which this reader does not parse. - Other users' files. Permissions decide readability, not the tool.
What it reports
Input, cached-input and output tokens, the model, the session and the date. Costs come from known API prices and are labelled estimates rather than subscription charges.
The cached-input line is worth keeping separate. Many trackers fold it into input, which overstates cost and hides a real optimisation: a long conversation that re-reads the same context is cheap or expensive depending on how the cache behaves.
What it deliberately does not report
Quota, balance and reset times. The guide says so, and the honesty is more useful than a guess would be.
That leaves a gap a team has to fill: the provider page. Checking one page once a week is a small price for not planning against an invented number.
Coverage beyond Antigravity
A single-tool tracker is easy to outgrow. UsageAtlas covers Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Gemini CLI, Amp, Qoder, Muse Code, Pi / OMP, Warp, Kimi Code, Kilo Code and Factory / Droid alongside Antigravity, all local-first with no account for local tracking.
The comparison across tools is the point. If one assistant consumes four times the tokens for similar output, that is a routing decision hiding in a monthly total.
How to evaluate any tracker
Ask four questions before installing anything.
- What format does it read, and does the tool actually write that format?
- Does it separate cached input from fresh input?
- Does it say what it cannot see?
- Can it run without an account and without reading prompts?
A tool that answers all four is worth the disk space. One that answers none is a dashboard with an opinion.
FAQ
Does UsageAtlas need an Antigravity API key?
No. It reads local history files, and the guide confirms no provider key is required.
Why does my history show nothing?
Either the CLI has not written compatible .db files, or the account running the tracker cannot read them.
Are the cost numbers accurate?
They are estimates from token counts and known API prices, not the amount a subscription charges.
The takeaway
Pick the tool that reads the real format and admits the limit. Skip the one that prints a percentage it cannot possibly know.