Antigravity Data Source (Experimental)

Antigravity support is experimental. ccusage can read Antigravity CLI usage from the per-conversation SQLite databases it writes locally, opening them read-only.

Read the accuracy notes before relying on these numbers. Antigravity stores usage as protobuf with no published schema, and some of its recording paths could not be verified against real data yet.

Quick Start

ccusage antigravity daily
ccusage antigravity monthly
ccusage antigravity session

Antigravity is also included in unified reports when supported databases are detected:

ccusage daily

Data Locations

By default, ccusage reads every conversation database under:

~/.gemini/antigravity-cli/conversations/**/*.db

Antigravity is a Gemini-family tool, so its CLI state lives under the shared .gemini directory rather than a directory of its own.

Set ANTIGRAVITY_DATA_DIR when Antigravity is stored somewhere else. It accepts comma-separated roots, so a current profile and an archive can be reported together:

ANTIGRAVITY_DATA_DIR="/path/to/antigravity-cli" ccusage antigravity daily
ANTIGRAVITY_DATA_DIR="/path/to/antigravity-cli,/archive/antigravity-cli" ccusage antigravity daily

With ANTIGRAVITY_DATA_DIR set, ccusage reads every .db file below each root's conversations/ directory, nested ones included:

<each-root>/conversations/**/*.db

Each conversation is a separate database, and a session in these reports is one conversation.

Token Mapping

Antigravity records one ModelUsageStats message per model invocation. ccusage reads:

Antigravity field	ccusage field
input_tokens	Input tokens
output_tokens	Output tokens
cache_write_tokens	Cache create tokens
cache_read_tokens	Cache read tokens
thinking_output_tokens + response_output_tokens	Output tokens, when the recorded total is absent
response_model	Model

input_tokens and cache_read_tokens are disjoint in the source, matching how ccusage reports them, so a cache read is never also counted as input. output_tokens already includes thinking tokens, so reasoning is not added on top of it.

Usage is read from two tables per conversation. steps is the primary source and holds one row per invocation, including background calls. gen_metadata is the only place a human-readable model name appears, and it also keeps usage visible for a step Antigravity has since pruned.

Cost Calculation

Antigravity leaves its model_cost, credit_cost and consumed_credits fields unset on consumer accounts, so there is no recorded cost to read. ccusage estimates cost from token counts and LiteLLM pricing for the recorded model name.

That figure is the equivalent public API cost, not an amount you were billed. Antigravity is quota-based, so the actual charge for this usage may well be zero.

--mode display therefore reports no cost, because there is nothing precomputed to display. Use --mode calculate to estimate cost from token counts instead:

ccusage antigravity daily --mode calculate

Use --offline to rely on cached pricing data:

ccusage antigravity daily --offline

The embedded offline snapshot covers Anthropic and OpenAI models. Gemini rates are only available from a runtime pricing fetch, so --offline reports no cost for Gemini models.

Accuracy Notes

Antigravity ships no .proto files, so ccusage decodes its stored blobs by protobuf field number. Those numbers were recovered from the schema descriptors embedded in the Antigravity binary rather than guessed from the data, and Antigravity has to read the same blobs back after an upgrade, so it cannot renumber a field without breaking its own persistence.

Within those constraints, the following limitations apply:

• Unnamed models cannot be priced. Antigravity records the model as a numeric id, and background invocations get no gen_metadata row to name them. These appear as antigravity-model-<id>, and ccusage warns that pricing is missing for them. Their tokens are still counted, so token totals stay complete while the cost total is understated.
• Cost is an estimate, never a billed amount. See Cost Calculation.
• Retries are handled by design but unverified. Failed attempts consume tokens and are recorded separately from the attempt that succeeded. ccusage collects both and deduplicates on the server-assigned response id, which is correct whether or not Antigravity repeats the successful attempt in its retry list — but this has not yet been checked against a conversation that actually retried.
• Subagents and model comparisons are unverified. Sub-conversations get their own database, so reading every database is what makes them visible, and the same response-id dedupe prevents double counting a call a parent also recorded. Neither path has been exercised against real data.
• Historical totals depend on retention. Antigravity trims what it keeps per conversation. If it prunes rows for an older day, that day's total will shrink; ccusage can only report what is still on disk.

Troubleshooting

If no Antigravity data appears, check that conversation databases exist under the default path or set ANTIGRAVITY_DATA_DIR.

ANTIGRAVITY_DATA_DIR="/path/to/antigravity-cli" ccusage antigravity session --json

Related Guides

• All Reports for unified multi-source reports
• Gemini CLI Data Source for the other Gemini-family source
• Environment Variables for the full list of path overrides
• Cost Modes for how --mode changes cost reporting