Antigravity Maestro

One extension. A pool of Google accounts. Three AI coding agents.
Run Antigravity's models in GitHub Copilot Chat, Claude Code and OpenAI Codex, from a pool that rotates itself when a quota runs out.

Installation · Quick start · Models · Gateway API · Configuration · Security

Overview

Antigravity Maestro turns the Google accounts you already have into a single, self-managing model pool. Sign in once per account; the extension tracks each one's quota, serves every request from whichever account can afford it, and exposes the whole pool to three clients at once — natively inside Copilot Chat, and over a local OpenAI/Anthropic-compatible gateway for Claude Code and Codex.

Not affiliated with, endorsed by, or sponsored by Google. "Antigravity" and "Gemini" are trademarks of Google LLC. The extension talks to the same endpoints the Antigravity client uses, authenticated with credentials you sign in with yourself.

Capabilities

Installation

From the VS Code Marketplace:

code --install-extension husamettinulutas.antigravity-maestro

Or search for Antigravity Maestro in the Extensions view. Requires VS Code 1.104 or newer.

Quick start

1. Open the Antigravity Maestro view in the activity bar.

2. Select Add account and complete Google sign-in in your browser. Repeat for each account — the first one becomes active.

3. Wire up the clients you use:

   Client	Action
   Copilot Chat	Set up on the Copilot Chat row, then Manage Models → Antigravity Maestro in the model picker, and tick the models you want.
   Claude Code	Use model on the Claude Code row (or run Antigravity Maestro: Use Model in Claude Code), then restart any running session.
   Codex	Use model on the Codex row. On Windows, restart VS Code once so Codex picks up the key.

Restore returns any agent to its own provider. Every config file is copied to <file>.antigravity-maestro-backup before the first write.

What the panel shows

The Accounts tab opens on Connections: whether the gateway is running, with its URL a click away, and which agents are wired to the pool. Below it, the account serving right now gets the whole stage, with a ring gauge per model family (Claude, Gemini, GPT-OSS) and when each one resets. If it is spending fast, the card says when it will run dry. If another account has more of that family left, the card offers to switch to it in one click. Under that, the 5-hour and weekly rolling windows. The rest of the pool is listed in rotation order, each account with its status in words, a pill per family, and Use to serve from it. Drag the list to set the order rotation falls back down. Refreshing, reordering and removing an account are in its ⋯ menu.

The Usage tab tracks what the pool actually spent, over the range you pick: this hour, today, yesterday, 7 days, 30 days or all of it. It starts with the total tokens served and the input / thinking / output mix; today and yesterday show each other for comparison, and the longer ranges add a bar per day. Then it breaks the spend down per model, each with its own input, thinking and output figures and the accounts that served it, and per account. Last, a quota history per account shows how fast each family is draining. Clear history… asks how much to remove: a recent span, everything older than a week or a month, or everything.

Models

Model ids are passed through exactly as upstream reports them — nothing is renamed. Where a model exposes separate reasoning tiers, those are distinct ids upstream with different thinking budgets, so they appear as distinct models here:

gemini-3.1-pro-high      gemini-3.5-flash-high      claude-opus-4-6-thinking
gemini-3.1-pro-low       gemini-3.5-flash-medium    claude-sonnet-4-6-thinking
gemini-3-flash           gemini-3.5-flash-low       gpt-oss-120b-medium

The authoritative list comes from each account's own quota response, so it reflects what Google actually serves that account.

The *-tiered Gemini Flash models are the exception: one id whose effort each request chooses. Copilot's model picker offers a Thinking Effort control for them where VS Code supports it — Low, Medium or High, about 1K or 4K tokens of thinking or the model's full budget — and antigravityMaestro.copilot.thinkingEffort sets the default. Claude Code's --effort and Codex's model_reasoning_effort set the same budgets, on every model that thinks, and the model is told the effort it runs at. Claude Code's ultracode reaches the model the same way: while it is on, the model is told to run substantive tasks as workflows, without being asked each time.

Each model in Copilot's picker shows the size of its context window, because the windows differ by an order of magnitude — around 1M tokens for the Gemini models against 200K for the Claude ones. Switching a long conversation to a smaller window makes Copilot Chat compact it first ("Compacting conversation…"), and that summary is a model turn like any other: it is charged to the account. Starting a new chat for the new model avoids it.

Gateway API

Claude Code and Codex speak HTTP, so the extension runs a server on 127.0.0.1:8765 (configurable), protected by a bearer key generated per install and held in SecretStorage. The three supported integrations are wired up for you by Use model — the gateway is there for everything else.

curl -s http://127.0.0.1:8765/health
curl -s http://127.0.0.1:8765/v1/models -H "authorization: Bearer $KEY"

Run Antigravity Maestro: Copy Local Gateway URL and Key to obtain $KEY, or use Copy URL + key on the gateway row to point a CLI, a script, or another editor at the pool. Restart rebinds the server after a port change, or if requests stop going through.

Copilot Chat does not use the gateway; those requests never leave the extension host.

Several windows, several accounts

Each VS Code window keeps its own serving account, and each project folder remembers the one it was on. Open two projects side by side, pick a different account in each, and neither moves the other. A window you have never picked in starts on the account you picked last.

Every window runs a gateway, and only one of them can hold the preferred port. The others start on the next free port but give agents the preferred address, so Claude Code and Codex configs always name 127.0.0.1:8765. When the window that holds the port closes, another window takes it over within seconds. Agents send their window's name with each request, so whichever window's gateway answers, the request runs on that window's account.

With the default highest-quota-first rotation, a window whose account runs out falls back to the next account in the list, which may be the one another window is using. Set antigravityMaestro.rotation.strategy to manual to keep every window on its own account.

Configuration

Under manual rotation the active account never changes on its own. Under the other strategies it changes only when the account you picked cannot serve the request.

Architecture

Copilot Chat ──► LanguageModelChatProvider ─┐
                                            ├─► account lease ──► Cloud Code endpoints
Claude Code ──► /v1/messages ──┐            │      (rotation,        (streamGenerateContent)
Codex ────────► /v1/responses ─┴► gateway ──┘       quota, tokens)

The account lease decides which signed-in account serves each request, refreshes its access token as needed, and records the token spend. Protocol translation is confined to src/protocol/, so a change to one wire format never touches the others.

Security

• Refresh tokens are stored in VS Code SecretStorage. They are never written to settings, logs, or workspace files.
• The gateway binds to 127.0.0.1 only and requires a bearer key generated per install.
• OAuth client credentials are injected at build time from a git-ignored .env and are never committed. See Building from source.
• Agent config files are backed up to <file>.antigravity-maestro-backup before any write, and Restore reverts them.

Report a vulnerability privately through GitHub Security Advisories.

Building from source

git clone https://github.com/husamettinulutas/antigravity-maestro.git
cd antigravity-maestro
npm install
cp .env.example .env      # then fill in your OAuth client

.env supplies AGM_OAUTH_CLIENT_ID and AGM_OAUTH_CLIENT_SECRET, which esbuild.js bakes into the bundle at build time. Building without it is supported — the result simply has no built-in client, and sign-in then requires the antigravityMaestro.oauth.* settings. Google issues the Cloud Code scopes only to approved clients, so the client you supply must be one of them.

Contributing

Issues and pull requests are welcome at github.com/husamettinulutas/antigravity-maestro. Run npm run compile && npm test before opening a pull request, and never commit credentials.

License

MIT © Hüsamettin Ulutaş