Antigravity Setup Guide

June 15, 2026 · View on GitHub

Track your Antigravity AI model quota usage with onWatch.

What is Antigravity?

Antigravity is an AI coding assistant that provides access to multiple AI models including Claude, Gemini, and GPT variants. onWatch can monitor your model quotas by connecting to the locally running Antigravity language server process.

Prerequisites

Antigravity installed and running
onWatch installed (Quick Start)

How It Works

onWatch automatically detects the Antigravity language server running on your machine by:

Scanning for the antigravity process
Extracting the CSRF token and port from command-line arguments
Connecting via the Connect RPC protocol
Polling the /exa.language_server_pb.LanguageServerService/GetUserStatus endpoint

No manual configuration is required for local development.

Data Sources: IDE vs CLI (agy)

Antigravity ships in more than one form, and all variants share a single Google-account quota, so onWatch shows one Antigravity card and lets you choose where the data comes from:

Source	What it does	Data richness
`ide`	Probes the running Antigravity desktop/IDE language server.	Per-model session quotas.
`cli`	Launches and manages the `agy` CLI in a pseudo-terminal and reads its `RetrieveUserQuotaSummary` endpoint.	Richer: weekly and 5-hour limits for the Gemini and Claude+GPT groups.
`both` (default)	Prefers the `agy` CLI for its richer data and falls back to the IDE probe when `agy` is unavailable.	Best available.

The dashboard card shows a "Source: agy CLI / IDE" badge indicating which probe produced the data you're seeing.

Choosing a source

During setup: onwatch setup asks which source to use when you enable Antigravity.
Anytime in the dashboard: open Settings -> Providers -> Antigravity and pick a Data Source. The change takes effect on the next poll - no restart needed.
Via environment: set ANTIGRAVITY_SOURCE=both|cli|ide in your .env.

About the `agy` CLI source

The agy CLI only exposes its quota server while an interactive process is alive, and it exits immediately without a terminal. To read it, onWatch launches a managed agy process inside a pseudo-terminal, keeps it warm between polls, relaunches it if it becomes unhealthy, tears it down when idle or on shutdown, and only ever stops the process it started - your own interactive agy sessions are never touched.

onWatch locates agy via ANTIGRAVITY_CLI_PATH, then your PATH, then common install locations (~/.local/bin/agy, /opt/homebrew/bin/agy, /usr/local/bin/agy). The CLI source is supported on macOS, Linux, and Windows (via ConPTY); the Unix path is the most thoroughly validated.

Quick Start (Auto-Detection)

Step 1: Enable Antigravity in onWatch

Add to your .env file:

cd ~/.onwatch

Edit .env and add:

ANTIGRAVITY_ENABLED=true

Or set it as an environment variable:

export ANTIGRAVITY_ENABLED=true

Step 2: Restart onWatch

onwatch stop
onwatch

Or in debug mode to verify:

onwatch --debug

You should see:

Antigravity auto-detection enabled (process scanning mode)
Starting Antigravity agent (interval: 60s)
connected to Antigravity language server port=63516 protocol=https

Step 3: View Dashboard

Open http://localhost:9211 and click the antigravity tab.

You'll see quota cards for each model, including:

Claude Sonnet
Gemini Pro
GPT-4 variants
And other available models

Docker/Container Configuration

In containerized environments where process scanning doesn't work, you can manually configure the connection:

Environment Variables

Variable	Description	Example
`ANTIGRAVITY_ENABLED`	Enable Antigravity provider	`true`
`ANTIGRAVITY_BASE_URL`	Base URL of the language server	`https://127.0.0.1:42100`
`ANTIGRAVITY_CSRF_TOKEN`	CSRF token from the process	`your_csrf_token_here`

Example Docker Compose

services:
  onwatch:
    image: ghcr.io/onllm-dev/onwatch:latest
    environment:
      - ANTIGRAVITY_ENABLED=true
      - ANTIGRAVITY_BASE_URL=https://host.docker.internal:42100
      - ANTIGRAVITY_CSRF_TOKEN=your_csrf_token
    ports:
      - "9211:9211"

Finding the Port and Token

On the host machine, find the running Antigravity process:

# macOS/Linux
ps aux | grep antigravity | grep -E "csrf_token|extension_server_port"

# Example output:
# /path/to/antigravity --csrf_token=abc123 --extension_server_port=42100

Extract the values:

--extension_server_port=42100 - Use as the port in ANTIGRAVITY_BASE_URL
--csrf_token=abc123 - Use as ANTIGRAVITY_CSRF_TOKEN

What Gets Tracked

Metric	Description
Model Quotas	Per-model remaining fraction (0.0 to 1.0)
Reset Times	When each model's quota resets
Prompt Credits	Available credits for your plan
Plan Info	Your subscription tier (Free, Pro, etc.)

The dashboard shows:

Usage percentage for each AI model
Remaining quota with color indicators
Reset countdown timers
Usage history and projections

Supported Models

onWatch tracks all models available in your Antigravity subscription:

Model ID	Display Name
`claude-4-5-sonnet`	Claude 4.5 Sonnet
`claude-4-5-sonnet-thinking`	Claude 4.5 Sonnet (Thinking)
`gemini-3-pro`	Gemini 3 Pro
`gemini-3-flash`	Gemini 3 Flash
(others)	Automatically detected

Troubleshooting

"Antigravity agent not starting"

Verify Antigravity is running:
```
ps aux | grep antigravity
```

Check if the process has the required arguments:

ps aux | grep antigravity | grep csrf_token

Ensure ANTIGRAVITY_ENABLED=true is set

"No models showing"

Make sure you're logged into Antigravity
Check that your subscription is active
Run onWatch in debug mode: onwatch --debug

"Connection refused"

The language server might be using a self-signed certificate. onWatch handles this automatically, but if you're using manual configuration:

Ensure the port is correct
Try both https:// and http:// protocols
Check firewall settings

"CSRF token invalid"

The token changes when Antigravity restarts. For auto-detection mode, restart onWatch after restarting Antigravity:

onwatch stop && onwatch

For manual configuration, update ANTIGRAVITY_CSRF_TOKEN with the new value.

API Details

onWatch uses the Connect RPC protocol to communicate with the Antigravity language server:

Endpoint:

POST /exa.language_server_pb.LanguageServerService/GetUserStatus

Headers:

Content-Type: application/json
Connect-Protocol-Version: 1
X-Codeium-Csrf-Token: <token>

Response structure:

{
  "userStatus": {
    "email": "user@example.com",
    "planStatus": {
      "availablePromptCredits": 500,
      "planInfo": {
        "planName": "Pro",
        "monthlyPromptCredits": 1000
      }
    },
    "cascadeModelConfigData": {
      "clientModelConfigs": [
        {
          "label": "Claude Sonnet",
          "modelOrAlias": {"model": "claude-4-5-sonnet"},
          "quotaInfo": {
            "remainingFraction": 0.75,
            "resetTime": "2026-02-24T12:00:00Z"
          }
        }
      ]
    }
  }
}

Security Notes

onWatch connects only to localhost (or your configured URL)
The CSRF token is never sent to external servers
All data stays on your machine (SQLite database)
Auto-detection only reads process arguments, not memory