Google Antigravity: Getting Started and Saving Money in Practice — Built-in Commands + Apprentice Asks the Master

Google's AI coding tool Antigravity has long kept a low profile, but thanks to advantages like no account bans, a rich model lineup, and cheap accounts on Xianyu, more and more developers are treating it as an alternative to Codex and Claude Code. A Bilibili creator, drawing on nearly a year of hands-on experience, breaks down the complete path from getting started to saving money: using built-in commands to "overclock" low-end models, using the "apprentice asks the master" method to squeeze every last drop of high-end web-based quota, and then connecting the desktop and web ends into a single pipeline via a self-developed browser extension.

What Antigravity Is and Why It's Worth Attention Now

Antigravity is Google's AI agent coding tool, positioned in the same category as Codex and Claude Code. In the video, the creator emphasizes that in terms of seniority, it is actually the earliest of this class of AI Agent products — it's just that Google's understated style has kept the product from standing out.

Two practical reasons have recently brought it back into view: Claude Code has been banning accounts frequently, and Codex quotas have dropped noticeably. Meanwhile, Antigravity accounts can be bought on Xianyu for just over ten yuan, and its model list includes both the latest Gemini models and mid-to-low-end models like Gemini 3.8 Flash — it can even access Claude Code 5.5, Opus, and Sonic. The creator himself uses a $20 account, whose quota structure is similar to the Xianyu student plan and the free mobile plan.

The quota panel is divided into three parts: Gemini 3.8 has its own section; Claude Code, Opus, and other models share one section with weekly and per-five-hour limits; and Gemini 4 has yet another separate section. The creator warns that high-end models like Gemini 4 and Opus burn through quota extremely fast — a $20 account is roughly enough for one day — while Gemini 3.8 Flash's quota is unbelievably large, lasting a week for non-heavy use.

Core Skill: Using the BOST Command to "Overclock" Low-End Models

This is the most important takeaway of the entire course. The creator says bluntly that Gemini 3.8 Flash is jokingly called the "American version of Doubao" — ask it a question directly and the accuracy may be only 30% to 60%, wrapping up in a minute. But after typing a slash to bring up the built-in command BOST, the same model performs completely differently: accuracy can reach around 90%, approaching first-tier model levels, and response time stretches out to planning very complex tasks — sometimes it can't even finish within an hour.

BOST essentially invokes multi-agent orchestration, breaking the problem down very finely and answering step by step. The creator has verified this with real projects: having the Flash model refactor software and develop games produced almost no errors once BOST was enabled. His advice is direct — when using low-end models, enable BOST without hesitation; it's the key to pulling a low-tier model up to top-tier quality.

Besides BOST, there are several other built-in commands worth remembering: learn makes the system reflect on and summarize recent successful experiences; grab me makes the AI act like an interviewer, asking questions back to align on requirements; and plan is used for planning at the project kickoff stage. Commands can be combined — for example, BOST first and then plan, boosting capability before laying out the plan. The creator specifically points out one difference between Antigravity and Codex/Claude Code: the latter two automatically find and invoke these strategies, whereas Antigravity requires the user to explicitly enter the commands.

Multi-Agent Orchestration refers to an architecture in which multiple AI agents with different roles or divisions of labor collaborate to complete a task: typically including roles such as planner, executor, and reviewer, which pass messages to break down goals, assign subtasks, and cross-verify results. Compared with single-turn Q&A, it can significantly reduce the error rate on complex tasks, at the cost of multiplied invocation counts and time. This also explains why the same low-end model's accuracy rises substantially after enabling BOST while runtime can stretch to an hour — essentially, a single rough answer is replaced by a multi-round pipeline of decomposition, execution, and review.

Apprentice Asks the Master: Letting the Web-Based High-End Model Work for the Desktop End

Overseas AI companies generally have two billing systems: web-based chat can't control the computer, while the desktop end that can code can see local files — and the quotas for the two are separate. This creates an arbitrage opportunity: when the desktop end's high-end model quota is used up, the web version of Gemini 4 often still has quota to spare.

The creator's approach is called "apprentice asks the master": the low-end model on the desktop end plays the apprentice, responsible for investigating local files, organizing requirements, and formulating good questions; the web version of Gemini 4 plays the master, responsible for planning, explaining, and proposing solutions — but it can't see the local situation, relying entirely on the apprentice to feed it context. Once the apprentice gets the answer, it is only responsible for execution, not decision-making. He offers an analogy: the chief engineer makes the plan, the workers do the work; the chief engineer can't see the site, so the workers must ask clearly about conditions on the ground.

The creator considers the web version of Gemini the most generous part: it can generate images, generate video with a daily free quota, create music, and offers Deep Research. For the Sanguosha card game he built, the background images, card assets, and background music were all generated by web-based Gemini, while the desktop low-end model handled slicing images, writing them to disk, and integrating them into the game. The whole project took three or four days, using only the $20 account.

Context refers to all the information a model can "see" when generating a response, including conversation history, uploaded file contents, retrieved materials, and system prompts. The model itself cannot access the user's local disk or runtime environment — its judgment about these depends entirely on what is placed into the context. Therefore, whether "apprentice asks the master" works hinges on whether the apprentice organizes the local file structure, error messages, requirement constraints, and so on into clear, complete, and not overly verbose context to pass to the master: missing information leads to solutions divorced from reality, while redundant information crowds the context window, raises costs, and dilutes the focus.

AgentLimb Extension: A Bridge Between Desktop and Browser

Linking the desktop and web ends requires a bridge. Half a year ago, the creator built his own Chrome-based browser extension, AgentLimb, which has been refined for six months and open-sourced, supporting Windows and macOS. The simplest way to install it is to hand the GitHub link to an AI and let the AI do it automatically; for manual installation, note that on Windows the extension directory must use an all-English path, and you need to enable Developer mode on the extensions management page before loading the unpacked package.

The extension's capabilities go beyond connecting to Gemini: it can operate the browser with multiple agents, multiple pages, and multiple accounts simultaneously, using different profiles to avoid repeated logins — the creator uses it to manage multiple Bilibili accounts and automatically upload videos. He mentions a real pain point: every AI company builds its own Chrome extension and they're not interoperable, so AgentLimb is positioned to use a single prompt to command all agents' browser operations in a unified way.

Browser Profile is an isolated environment within Chrome and browsers on the same engine, containing separate cookies, login states, local storage, and extension configurations. By launching multiple browser instances with different profiles, you can keep multiple accounts logged in simultaneously on the same machine without interference — this is also the foundation for multi-account automation. Note that the terms of service of most platforms explicitly restrict automated multi-account operations; you should assess compliance and ban risks yourself in actual use.

Live Demo and the Money-Saving Logic

In the video, the creator demonstrates the full workflow: specify a desktop folder, have the Flash model take a "test the game" requirement to ask the web-based master, the master returns an MVP plan for Snake or Sanguosha, and the apprentice automatically writes it to disk and starts development. In between, the model automatically confirms the connection, takes screenshots to assess the page, and organizes questions — no manual copy-pasting throughout. He deliberately left BOST off to speed up the demo, noting that it should be enabled when seriously working on a project.

The money-saving logic boils down to two points: first, use BOST to pull Flash up to near top-tier model quality, reducing reliance on expensive models; second, use "apprentice asks the master" to squeeze the web quota dry as well — effectively getting two quotas' worth from one account. The creator's attitude is pragmatic — if money is no object, go straight to the Claude Code commercial edition; if the budget is limited, use this method to achieve a close result. He also reminds people that for any AI project, first have a high-end model write the plan and do a minimal viable MVP validation, then let the low-end model execute — don't start building a skyscraper right away.

One-line Summary

Use the BOST command and the "apprentice asks the master" method to get top-tier results from Antigravity's low-end models and squeeze two quotas dry.

Paragraph Summary

Based on a Bilibili creator's nearly a year of hands-on experience, this article systematically breaks down the getting-started and money-saving path for Google's AI coding tool Antigravity. Thanks to no account bans, a rich model lineup, and cheap Xianyu accounts, it is becoming an alternative to Codex and Claude Code. There are two core techniques: first, the built-in command BOST, which through multi-agent orchestration raises Gemini 3.8 Flash's accuracy from 30%–60% to around 90%, approaching first-tier models; second, the "apprentice asks the master" method, which exploits the separation of web and desktop quotas — letting the desktop low-end model organize context and ask questions while web-based Gemini 4 handles planning — thereby getting two quotas' worth from one account. The accompanying open-source extension AgentLimb bridges the desktop and browser, supporting multi-agent, multi-account, and multi-page operations to form a complete pipeline.