Antigravity Tier

antigravity-tier

Antigravity's free Individual-tier quota is a real second compute tier, reachable headlessly:

agy -p "<self-contained prompt>" --sandbox --new-project

The flag parser is strict since agy 1.1.18 — the prompt must sit immediately after -p (or use -p='<prompt>'); agy -p --sandbox '<prompt>' errors with "-p took --sandbox as its prompt". For a long job (deep research, a multi-page draft) raise --print-timeout past its 5-minute default, and pass --effort high when the work needs real reasoning depth. --model <slug> / the /model slash command pick a specific model (see roster below) and persist it as default.

Confirmed 2026-08-20: works, authenticated, returns real output for genuinely isolated prompts. Always pass --sandbox — real, confirmed CLI flag ("run in a sandbox with terminal restrictions enabled"), unlike a persistent settings.json which was never verified to exist or apply for this install. Never pass --dangerously-skip-permissions from a scripted or ad-hoc call — the one sanctioned exception is run-image-bridge.ps1, which scopes it narrowly to the reference-image-read case only.

--new-project is required, not optional, for any call that should see a project's own GEMINI.md. Confirmed 2026-08-26: plain agy -p from a project directory does NOT bind that directory as the workspace root — agy reports "no active workspace" and falls back to default-cli-project (an empty bucket), loading only the global ~/.gemini/GEMINI.md. --new-project is what actually binds cwd as the root the rules walk-up starts from. Skills, by contrast, are global and load regardless of this flag.

What Antigravity can now do — the ported skill set (2026-08-26)

Beyond raw model calls, agy now carries its own skill roster at ~/.gemini/config/skills/ (global, same open SKILL.md standard, no --new-project needed to see them): gfx-prompt (Roblox GFX render-prompt craft), image-handoff (package a prompt for a web tool instead of generating here), image-bridge-loop (request→render→verify discipline; on image-quota exhaustion agy now emits its own gemini.google.com / ChatGPT / Midjourney handoff block and writes the prompt to brain/<uuid>/scratch/<name>_prompt.txt, overlapping image-handoff — still watch for the older silent fallback, a flat vector shape masquerading as a real render), humanizer, sequential-thinking, agy-research (web research via the exa MCP already registered for agy), and local-tier. A job that needs one of these no longer has to carry the whole method in the prompt — mentioning the task is often enough for agy to reach for the skill itself.

Project rules exist for ClaudeFX/Projects/SoundWave_OG/GEMINI.md and ClaudeFX/Projects/Toy_Obby/GEMINI.md — real project paths, not the C:\Dev\SoundWave_OG\GEMINI.md path once assumed and confirmed absent.

Which brain does which job

Negative rules that hold across all three: no MCP servers or skills reach a job routed through the LM Studio API — only the model and the preset do. No BYOK on Antigravity — it cannot be pointed at LM Studio or a personal Anthropic key. Never register the roblox-studio MCP inside a bridged LM Studio session — its schema crashes the grammar compiler outright.

When to use this, not Claude Code

Only when the job needs zero of Claude Code's own tooling: no MCP servers, no subagents, no vault memory, no file edits that need review. If the job needs any of that, it stays on Claude Code — free output that's wrong or unreviewed is not actually cheap.

Quota — unverified, don't assert a number

Google publishes no exact free-tier limit ("meaningful quota, refreshed weekly" is the only language in their docs). A /usage prompt sent via -p does not return a quota readout — it makes a read_file tool call that headless mode auto-denies ("a tool required the read_file permission that headless mode cannot prompt for"). Same net result: no scripted quota visibility, so do not tell the user this shows remaining quota. Just make the call; if it's quota-exhausted, agy returns a 429 Resource Exhausted error (and, for an image job, its own web-tool handoff block), not a wrong-but-confident answer.

Output handling

Output lands in the CLI's own text response — unlike image generation (see image-request skill), there is no separate file to hunt for. Write the result to a file yourself if it needs to persist, and treat it as another agent's output: review before passing it on, the same as any subagent's report.