Antigravity

The antigravity adapter drives the Antigravity CLI in its headless stdout mode, detected as the agy binary on PATH; import antigravity from anyagent-js/antigravity.

What you can ask for

model, effort (closed vocabulary: low, medium, high), and resume work on the CLI’s own flags; systemPrompt and schema are provided by AnyAgent. readOnly and mcp are undeclared and throw: no true read-only run exists, since --mode plan still leaves Antigravity’s internal scratch directories writable. Plan mode stays reachable via extraArgs: ["--mode", "plan"]. models() runs agy models, a cross-vendor catalog whose ids are valid verbatim as model and can bake in an effort level; authStatus() is probed, reading a model listing as authenticated. usageStatus() reads the account-wide weekly quota. The full grid is on the support matrix.

Good to know

• Stdout mode times out after 5 minutes by default; give a longer run extraArgs: ["--print-timeout", "30m"].
• The init event’s conversation id lands on result.sessionId and resumes via resume.
• usage maps the CLI’s thinking_tokens to reasoningTokens; no cache line items are reported.
• usageStatus() needs an Antigravity product running — the desktop app, the IDE, or an interactive agy session — because the quota is served by their local language server. With none up it answers { state: "unknown" }.
• That server is on loopback, so this needs no network opt-in: nothing leaves the machine. Quota is account-wide, so whichever product answers first is authoritative.
• Buckets are per model group, not per model: { model: "claude-opus-4-6-thinking" } narrows to the Claude/GPT pool, gemini-* to the Gemini one. Which buckets exist varies by plan — do not assume a 5-hour one is present.

Under the hood

Every run executes agy -p <prompt> --output-format stream-json --dangerously-skip-permissions, with the prompt as -p’s value; piped stdin is ignored in stdout mode. Run options append flags: