supergravity-mcp

An MCP server that lets Claude Code (or any MCP client) delegate tasks to the Google Antigravity desktop app — by driving its real, already-logged-in UI. No reverse engineering, no extracted credentials, no unofficial API calls.

Why this exists

Antigravity ships a CLI (agy) for scripting it headlessly — but on older Intel Macs, that CLI isn't available even though the desktop app runs fine (confirmed working on a 2015 MacBook Pro, macOS Monterey, Intel). This project is the missing bridge for anyone in that situation: it automates the app the same way a human would use it, so it works anywhere the app itself works.

How it works

1. Antigravity is relaunched with Chrome's remote-debugging flag (--remote-debugging-port), the same mechanism browser test tools like Playwright use to drive Electron apps.
2. A script connects to that debug port and: types your task into the chat box (via a simulated paste event — Antigravity's input is a rich-text editor, not a plain textarea, so it needs a real paste, not just poking text into the DOM), clicks send, and reads back the reply once it's done streaming.
3. Everything happens through your own existing login. This tool never touches Antigravity's stored credentials, internal code, or private APIs.

Setup

Antigravity must be installed and already signed in. By default the tool looks for it at /Applications/Antigravity.app or ~/Desktop/Antigravity.app; set ANTIGRAVITY_APP_PATH if it lives somewhere else.

Easiest: install the plugin (MCP server + skill together)

claude plugin marketplace add presidentrice/supergravity-mcp
claude plugin install supergravity@supergravity-mcp

This gets you both the delegate_to_antigravity / get_antigravity_quota tools and the supergravity:delegate skill, which teaches Claude how to route natural requests ("use gemini", "ask antigravity", "check antigravity's quota") to the right tool call and model.

Or: just the MCP server, no skill

claude mcp add supergravity -- npx -y supergravity-mcp

Works fine without the skill — Claude still has the tools, just less built-in guidance on picking a model or handling a quota-exhausted error.

Building from source

npm install
npm run build

The tools

delegate_to_antigravity(task, model?, timeoutMs?)

• task — what to ask Antigravity to do.
• model — optional. One of the models in Antigravity's own selector (see AVAILABLE_MODELS in src/antigravity-client.ts).
• timeoutMs — optional, default 120000 (2 minutes).

If the task generates files (images, documents, etc.), the result also includes a real filesystem path — Antigravity stores each conversation's files at ~/.gemini/antigravity/brain/<conversation-id>/ on disk, confirmed directly (not documented anywhere by Google). No need to fetch anything through Antigravity's local server; just read the file.

delegate_to_antigravity_batch(tasks)

Runs several independent tasks at once, each in its own Antigravity window — genuinely concurrent, not queued. Confirmed with two simultaneous ~200-word generations that both completed correctly in the same ~16s window. Opens extra windows automatically as needed (macOS's "New Window" menu command via System Events, with a fallback for when zero windows are open).

Requires Accessibility permission for whatever process runs this (System Settings > Privacy & Security > Accessibility) — macOS only. Don't run this at the same time as a separate delegate_to_antigravity call; window allocation between the two isn't coordinated.

delegateToAntigravityBatch([
  { task: "..." },
  { task: "...", model: "Claude Sonnet 4.6 (Thinking)" },
]);

list_antigravity_files(limit?)

Every file Antigravity has generated, across every conversation, most recent first — not scoped to a single delegate call. Useful for finding something generated earlier in this session or a past one.

get_antigravity_quota()

Reads Antigravity's Settings > Models panel and returns remaining quota — Gemini models and Claude+GPT models draw from separate pools, each with a weekly limit and a 5-hour limit. Useful to check before delegating, or to explain a send button is disabled error (usually means that pool is exhausted, not that the model is unavailable).

Picking a model

Antigravity doesn't show per-message token cost, only pooled quota %. The model parameter's description carries a rough cost/capability cheat sheet (MODEL_GUIDE in src/antigravity-client.ts) built from third-party API list prices and public benchmarks, researched mid-2026 — newer than this tool's own training data, so treat it as a starting point, not gospel:

Model	~Price (in/out per 1M tokens)	Known for
Gemini 3.5 Flash	$1.50 / $9	Fastest, cheapest frontier-tier option; strong coding/agentic benchmarks
Gemini 3.1 Pro	$2 / $12	"Thinking" mode for harder reasoning (ARC-AGI-2 ~77%)
Claude Sonnet 4.6	$3 / $15	Coding quality, computer-use, agent workflows (SWE-bench ~80%)
Claude Opus 4.6	$5 / $25	Strongest complex/ambiguous reasoning; legal/financial/research-grade analysis
GPT-OSS 120B	~$0.03–0.09 / $0.10–0.40	Open-weight, near-o4-mini reasoning, far cheaper than the rest

These are what the underlying models cost via their own APIs elsewhere — Antigravity itself doesn't bill you per token, it draws down the pooled quota above instead.

Known limitations

• A model's quota can run out. If a non-default model (e.g. Claude Opus) has hit its usage limit, Antigravity's send button stays disabled after selecting it, even though it's still pickable in the dropdown. The tool fails fast with a clear error in that case instead of hanging until timeout — just retry later or pick a different model.
• UI automation, not an API. If Google changes Antigravity's screen layout, the CSS selectors this relies on (data-testid="send-button", the message input's aria-label, etc.) may need updating. That's the tradeoff for staying entirely inside Antigravity's terms of use — see below.
• Reply extraction is a text heuristic, not a structured read. It looks for the "Thought for Ns" marker Antigravity renders before its answer and takes what follows. Works reliably in testing but is the most likely thing to break on a UI redesign.
• Single-window assumption. Only tested with one Antigravity window open.

What this deliberately does NOT do

No unpacking of Antigravity's app.asar, no reading of stored auth tokens, no calling of internal/undocumented endpoints. Everything here operates through the same interface a human uses, which is why it doesn't carry the legal risk that reverse-engineering approaches do.

License

MIT