Build Notes: Getting Started with Antigravity & The Operator's Lab

In AGY-native development, you don’t sit in an editor manually writing boilerplate code, configuring database schemas, or hand-crafting shell scripts. Code, configuration files, and terminal setup commands are disposable implementation details managed entirely by Google Antigravity (agy).

What actually matters is the architecture, the prompting philosophy, the security perimeter, and how you direct the agent.

If you want to build autonomous agents that touch real data, physical hardware, and financial systems without burning money or leaking private information, this is the foundational blueprint.

What is Google Antigravity? (CLI vs. IDE)

Google Antigravity is Google’s agentic development ecosystem, designed to transform how software engineering, data pipeline construction, and system administration are performed. Antigravity exists in two primary formats:

Antigravity IDE: A rich graphical desktop development environment (GUI). While full-featured, running a heavy desktop editor window consumes 2GB+ of system memory—taxing consumer hardware when paired with local AI runtimes.
Antigravity CLI (antigravity-cli or agy): A lightweight Command Line Interface (CLI) tool that operates directly inside your terminal window. This is what we recommend and use throughout this entire series.

How to Install `antigravity-cli`

Installing antigravity-cli takes less than two minutes. The official documentation provides simple setup scripts for macOS (via Homebrew), Windows, and Linux:

📥 Official Setup Instructions: Install Antigravity CLI (https://antigravity.google/product/antigravity-cli)

Defining the CLI & How Cloud Intelligence Interacts with Local Files

Operating via a CLI (Command Line Interface) means interacting with software by typing plain-text instructions into your terminal application, running natively on your local machine.

It is important to understand how antigravity-cli balances cloud intelligence with local execution:

Cloud Reasoning (Google Gemini): Antigravity connects to Google Gemini (a high-reasoning cloud LLM) to analyze complex goals, plan multi-file architectures, and debug logic flaws. Your prompts and code queries are processed securely by Gemini in the cloud.
Local On-Device Authority: Unlike a standard browser chat window (like ChatGPT or Gemini in a web tab) that is completely isolated from your computer, antigravity-cli runs locally on your machine. This gives the agent the local authority to:
- Read and write local files: Author Markdown context files in knowledge/, write Python analytical scripts, or generate SQL queries directly on your hard drive.
- Execute local terminal commands: Run compilation tools, run local databases (like SQLite or BigQuery local tools), and flash microcontroller firmware (ESP32).
- Build persistent Skills & tools: Codify reusable prompt workflows and Model Context Protocol (MCP) integrations on-device.

Terminal Choice: Default Terminals vs. Upgraded (Ghostty)

Both macOS and Windows ship with built-in default terminal applications out of the box (Terminal.app on Mac; Command Prompt / PowerShell / Windows Terminal on Windows).

While default terminals work perfectly fine to get started, we strongly suggest upgrading to a modern, high-performance terminal emulator like Ghostty (or iTerm2 on macOS / Windows Terminal on Windows):

Minimal Memory Tax: antigravity-cli running inside Ghostty uses a tiny ~50MB RAM footprint, leaving 6GB+ of memory free on an 8GB M2 Mac Mini to run local models without swap thrashing.
Speed & Ergonomics: Modern terminals provide GPU acceleration, sub-5ms input latency, crisp font rendering, and effortless multi-pane window management.

The Prompting Philosophy: Act as Product Lead, Not Keyboard Typist

The biggest mental shift in AGY-native development is moving from how to code it to what business result needs to be achieved.

Stop trying to write prompt templates that guess SQL syntax or Python regex. Act as the Product Lead and Architect: define plain-English intent, specify non-negotiable constraints, enforce security boundaries, and let agy author and execute the implementation.

The Ground-Truth Knowledge Base Pattern (`knowledge/`)

Instead of dumping unstructured background into every prompt or relying on complex vector databases, maintain plain-text Markdown files in a dedicated knowledge/ directory:

knowledge/accounts_and_entities.md: Living trust catalog, custodians, checking accounts, and private equity holdings.
knowledge/hardware_and_memory.md: RAM allocations, terminal footprint targets, and headless host rules.
knowledge/network_security_rules.md: Firewall default-deny policies, Tailscale SSH mesh rules, and identity gating.

When plain text files hold your ground truth, both you and agy can inspect, audit, and edit rules with zero database re-indexing overhead.

The Pragmatic 4-Step Getting Started Blueprint

Here is the exact sequence of steps to establish your own operator lab setup from scratch.

Step 1: Terminal & Remote Session Persistence Setup

Do not run heavy GUI IDEs alongside local AI models on consumer hardware. Run antigravity-cli headlessly inside a lightweight terminal like Ghostty or tmux.

Maintain an always-on session inside tmux so long-running agent builds survive network disconnects when working remotely.
🔗 For setting up zero-trust remote access, hardware YubiKey identity gating, and persistent 1-tap mobile sessions from an iPad or iPhone, see Build Notes: The Remote Control Plane.

Step 2: Zero-Trust Security Perimeter & On-Device Scrubbing

Never expose public ports or send raw financial statements to un-encrypted cloud APIs.

Zero Public Open Ports: Gate all host access behind Tailscale SSH and hardware security key identity (YubiKey / Google Advanced Protection).
On-Device PII Scrubbing: Run lightweight local models (qwen3.5:4b via Ollama) to parse PDFs and mask sensitive PII (SSNs, 16-digit card numbers into XXXX-1234) on-device before any payload touches cloud storage.
🔗 For setting up private, on-device biometric parsing without data leaving your machine, see Build Notes: The Zero-Cloud Fitness Coach.

Step 3: GCP Cost Guardrails & Order-of-Magnitude Safeguards

Cloud data warehouses like Google BigQuery bill by bytes scanned. Prevent runaway query bills with automated guardrails:

Dry-Run Scan Checks: Direct agy to run DRY_RUN checks to estimate totalBytesProcessed before executing any SQL query.
Order-of-Magnitude Sanity Gates: Embed Python verification scripts that sanity-check output balances before updating production ledgers.
🔗 For configuring dry-run query scan checks and auditing multi-institution financial fee drag, see Build Notes: The Private Family CFO.

Step 4: Multi-Model Routing & Adversarial Peer Review

Do not send every task to the largest cloud model. Match each job to the cheapest capable model:

Tier 1 (Cloud Orchestrator - Gemini): High-level architectural planning and multi-file code execution via agy.
Tier 2 (Adversarial Reviewer - Codex CLI): Spawns an independent model family in a subagent shell to stress-test draft code before committing changes.
Tier 3 (Local Execution Engine - Ollama): Runs local 4B models for zero-cost background tasks and private text parsing.
🔗 For setting up this 3-tier multi-model routing topology on an entry-level 8GB Mac Mini, see Build Notes: Multi-Model on Cheap Hardware.
🔗 For controlling physical hardware and microcontrollers via local Home Assistant loops, see Build Notes: The Thermal Battery.

Summary of the Operator Lab Setup Blueprint

Layer	Recommended Tool / Pattern	Role & Purpose
Orchestrator	`antigravity-cli` (`agy`)	Autonomous AI Financial Engineer & System Architect
Ground Truth	`knowledge/*.md` files	Living catalog of rules, account specs, & hardware limits
Terminal Host	Ghostty + `tmux`	Headless execution preserving 6GB+ RAM for local LLMs
Zero-Trust Mesh	Tailscale SSH	Hardware key-gated access with zero public open ports
On-Device LLM	Ollama (`qwen3.5:4b`)	Zero-cost local PII scrubbing & private file processing
Cloud Warehouse	Google BigQuery	Relational storage for multi-year cash flow & fee audits
Cost Guardrails	GCP `DRY_RUN` checks	Pre-query byte-scan estimation keeping cloud bills to pennies

Operating Your Lab Day-to-Day

Once agy deploys this foundational stack, building new capability is fast and cost-effective:

Define Intent: State what problem you want to solve in plain English.
Review Proposed Plan: Let agy evaluate the architecture and propose a multi-step execution plan.
Approve Execution: Grant permissions for shell execution, file creation, or GCP tool invocations step-by-step.
Inspect & Commit: Verify outputs using interactive CLI tools and live staging previews.

By acting as the System Architect while agy handles code generation, cloud schema creation, and system configuration behind the scenes, you build production-grade personal systems for just $40/month in total marginal cost.

📖 Companion Article

Read the main narrative story: An Operator's AI Lab: Series Overview.

Comments welcomed:

Discuss on X / Twitter → Discuss on LinkedIn →