Antigravity

   ____ _                 _
  / ___| | __ _ _   _  __| | ___
 | |   | |/ _` | | | |/ _` |/ _ \
 | |___| | (_| | |_| | (_| |  __/
  \____|_|\__,_|\__,_|\__,_|\___|

               |
          +----+----+
          |   MCP   | (- you are here
          +----+----+
               |

     _          _   _                       _ _
    / \   _ __ | |_(_) __ _ _ __ __ ___   _(_) |_ _   _
   / _ \ | '_ \| __| |/ _` | '__/ _` \ \ / / | __| | | |
  / ___ \| | | | |_| | (_| | | | (_| |\ V /| | |_| |_| |
 /_/   \_\_| |_|\__|_|\__, |_|  \__,_| \_/ |_|\__|\__, |
                      |___/                       |___/

What is this?

mcp-server-google-antigravity is a small MCP (Model Context Protocol) server. It exposes Google Antigravity — Google's terminal coding agent, driven by the agy CLI (Gemini under the hood) — as a set of tools any MCP client can call.

The idea: your primary model (Claude Desktop, Claude Code, Codex, …) stays lean and delegates the grunt work — web search, whole-repo analysis, scaffolding files, long-running edits — to a second agent running in the background. Jobs are async and unlimited: you fire a task, get a jobId instantly, and poll for the result while the caller keeps working.

   Claude / Codex          this MCP server            Antigravity
  +---------------+  tool  +----------------+ spawns  +--------------+
  |  "analyze the | -----> | use_antigravity| ------> | agy  --print |
  |   whole repo" |        | -> jobId(async)|         | (Gemini)     |
  +---------------+ <----- +----------------+ <------ +--------------+
        keeps working        polls result / logs        does the work

Requirements

Step 1 — Install the Antigravity CLI (agy)

Never used agy before? Start here. It's one command. First launch signs you in with Google.

macOS / Linux

curl -fsSL https://antigravity.google/cli/install.sh | bash

Windows (PowerShell)

irm https://antigravity.google/cli/install.ps1 | iex

Windows (CMD)

curl -fsSL https://antigravity.google/cli/install.cmd -o install.cmd && install.cmd && del install.cmd

Where it lands (remember this — you may need it for AGY_PATH below):

Sign in and verify:

agy            # first run opens a Google Sign-In in your browser (creds cached in your OS keyring)
agy --version  # should print a version, e.g. 1.1.1
agy models     # lists the models you can use (Gemini 3.5 Flash, Gemini 3.1 Pro, ...)

If agy isn't found afterwards, open a new terminal (so PATH reloads) or run agy install to fix shell paths.

Step 2 — Get this MCP server

Option A — npx (recommended): Skip straight to Step 3 and use the npx config. The npx path works seamlessly from the public npm registry with no authentication required.

Alternatively, you can install the package globally:

npm install -g mcp-server-google-antigravity

This exposes the mcp-server-google-antigravity command (which you can register directly in your MCP client's config instead of npx).

Option B — from source (for hacking on it):

git clone https://github.com/TurkerYakup/mcp-server-google-antigravity.git
cd mcp-server-google-antigravity
npm install
npm run check   # quick syntax check

node-pty (used only for clean output on Windows) is an optional dependency — if it can't build on macOS/Linux the install still succeeds and the server falls back automatically. No compiler needed.

Step 3 — Register it with your MCP client

{
  "mcpServers": {
    "antigravity": {
      "command": "npx",
      "args": ["-y", "mcp-server-google-antigravity"],
      "env": {
        "AGY_MODEL": "Gemini 3.5 Flash (Medium)",
        "AGY_AUTO_APPROVE": "true"
      },
      "timeout": 900000
    }
  }
}

# npx (Option A)
claude mcp add antigravity -- npx -y mcp-server-google-antigravity

# from source (Option B)
claude mcp add antigravity -- node /absolute/path/to/mcp-server-google-antigravity/index.js

That's it. Your client now has tools like use_antigravity, antigravity_continue, and the filesystem helpers.

Quick example

Ask your MCP client something like "use antigravity to analyze the repo at C:/Dev/Repos/foo and summarize its architecture". Under the hood the client calls:

1. use_antigravity({ prompt: "...", add_dirs: ["C:/Dev/Repos/foo"] }) → returns { "jobId": "mrib1q20_wa5ma1", "status": "running" } instantly.
2. antigravity_result({ jobId, wait_ms: 10000 }) → blocks/long-polls until the job is done or the 10-second wait elapses. While still running, it returns a lean payload containing status, character count, and the last 200 characters of output ({ status, jobId, chars, tail, startedAt }). Once complete, it returns the full record (including durationMs and bytes telemetry).

Follow up in the same session with antigravity_continue({ prompt: "now list the top 3 risks" }). Need agy to write files? Pass add_dirs so it has a workspace; auto_approve (default) lets it create/edit without prompts.

Configuration (environment variables)

Everything is optional — sane defaults out of the box.

Tools

The server-side filesystem tools use Node's fs directly — instant, and never gated by a permission prompt.

Live progress notifications

Because jobs return a jobId immediately, the server pushes MCP logging notifications (notifications/message, logger antigravity) so the client can see activity without polling:

• job started — jobId, kind, prompt snippet
• job running — heartbeat every AGY_HEARTBEAT_MS (default 15s) with a tail of the current partial output
• job done / job failed — final status, byte count / error

Best-effort: if the client didn't negotiate the logging capability they're silently dropped. Polling antigravity_result works either way.

How it works (design notes)

The agy CLI runs a single prompt with agy --print "<prompt>" and returns when done. Three problems and their fixes:

1. MCP client request timeout (~60s). Long tasks made the client give up even though agy kept running. → Async job pattern: use_antigravity spawns agy in the background, returns a jobId instantly, writes the result to a JSON file in the OS temp dir (agy_jobs/), and the client polls antigravity_result. Optional long-polling (wait_ms) lets clients wait inside the request when short runs are expected. No request ever blocks.
2. agy's own print timeout (default 5m). Heavy prompts hit timeout waiting for response. → every job passes --print-timeout 10m.
3. Headless file-permission gate. In --print mode agy's file-writes get denied by the "ask" policy and the job hangs. → auto_approve (default true) passes --dangerously-skip-permissions; add_dirs passes --add-dir <path> for the workspace it needs.

Security note

auto_approve defaults to true (--dangerously-skip-permissions) — agy will create/edit files and run terminal commands without prompting. This is intentional: in headless --print mode agy otherwise blocks on a permission prompt and the job hangs. For untrusted prompts, prefer sandbox: true (terminal-restricted, cheap — a permission layer, not a VM) over turning auto-approve off. Use auto_approve: false only for genuinely read-only asks.

Platform support

Fully cross-platform — macOS, Windows, and Linux. Windows gets a node-pty path for clean streaming (with an automatic fallback if it's unavailable) and auto 8.3 short-path resolution for non-ASCII usernames; macOS/Linux use plain child_process. The only per-OS difference is where agy lives and where your client's config file is (both covered above).

Troubleshooting

License

MIT © Türker Yakup Altınsoy.