---
title: 'Antigravity Agents That Run 100% Offline (FREE) 🤯'
source: 'https://youtube.com/watch?v=1Fq87QKf17Q'
video_id: '1Fq87QKf17Q'
date: 2026-11-17
duration_sec: 656
channel: 'Julian Goldie SEO'
---

# Antigravity Agents That Run 100% Offline (FREE) 🤯

> Source: [Antigravity Agents That Run 100% Offline (FREE) 🤯](https://youtube.com/watch?v=1Fq87QKf17Q)

## Summary

The video demonstrates how to build AI agents using the Anti-Gravity SDK that run entirely offline on a local machine, powered by models like Gemma 4 via Ollama. The creator showcases four practical agents—Ivy for client email triage, Leo for prospect outreach, Nova for session notes, and Sol for SEO dashboards—all operating without internet, API keys, or cloud data exposure.

### Key Points

- **Introduction to Local AI Agents** [00:02] — Anti-Gravity agents can now run on your own computer with a local model, no internet needed. Four agents are introduced: Ivy (client emails), Leo (prospect openers), Nova (coaching notes), and Sol (SEO dashboard).
- **Google's Announcement** [00:45] — Google announced local model support for the Anti-Gravity SDK. The headline setup uses Gemma 4 (26B parameters) requiring at least 24GB memory, but the SDK also works with any OpenAI-compatible local server like Ollama, LM Studio, or VLLM.
- **Hybrid Mode** [01:41] — Google demonstrated a hybrid mode where a cloud model plans the job (seeing only file names and task descriptions) while the local model does the actual work. In their demo, 97.2% of tokens ran locally.
- **Community Demand** [02:10] — Members are asking for agents that run all day on their own files without worrying about keys, limits, or data privacy. Examples include local document search and vision models for video footage.
- **Setup Steps** [03:06] — Step 1: Install Ollama and use Gemma 4 E4B (runs on 16GB). Step 2: Open Anti-Gravity and start a new conversation. Step 3: Ask Anti-Gravity to build the agent using the SDK with local OpenAI config, pointing to Ollama.
- **Prompt Template** [04:03] — The exact prompt includes: 'Build a Python agent with the Anti-Gravity SDK. Use local OpenAI agent config with the model Gemma 4 E 4 B and the base URL for Ollama on localhost. Set the workspace to this folder. The agent reads every file in the input folder, does the job I describe, and writes its results to the output folder. Print every tool call so I can watch it work. Then run it.'
- **Ivy: Client Inbox Agent** [04:29] — Ivy reads client emails, ranks them by priority, writes summaries, and drafts replies. It runs locally, keeping client data private. Example: it flagged an unhappy client email as high priority and drafted a calm reply.
- **Leo: Prospect Opener Agent** [05:41] — Leo writes personalized first lines for leads based on their business details. For a course creator with 4,000 students, it mentions their busy help inbox; for a SaaS tool, it talks about guiding trial users.
- **Nova: Session Notes Agent** [06:38] — Nova turns messy coaching notes into clean plans with three action items and one question for the next session. It handles private data locally, which is crucial for coaching notes.
- **Sol: SEO Dashboard Agent** [07:23] — Sol converts a search console export into a single HTML dashboard with top queries, quick wins (keywords between positions 8-20), and a bar chart. Ideal for SEO agencies to hand to clients monthly.
- **Safety Features** [08:10] — Policies allow blocking tools by default, requiring permission for commands. Triggers let agents check folders periodically (e.g., every 10 minutes). Custom tools and MCP servers can be added.
- **Mistakes to Avoid** [08:54] — 1) Always give the agent its own workspace folder. 2) Give one clear job with exact file names. 3) Provide room to think—Google recommends 64,000 token context; raise context size in Ollama if jobs stop halfway.
- **Recommendations** [09:53] — Agencies should start with Ivy, automation builders with Leo, coaches with Nova, and SEO agencies with Sol. Course owners can copy Nova's pattern for student questions.
- **Honest Take** [10:21] — Small local models are slower than cloud models and may need checking on harder jobs, but they're brilliant for repeat jobs on private files and run as often as you like.

### Conclusion

Local AI agents offer a privacy-preserving, cost-effective solution for repetitive tasks on sensitive data. By following the simple setup and avoiding common mistakes, anyone can deploy these agents to streamline their workflow.

## Transcript

Anti-Gravity, [music] and it changes how you build agents. Your Anti-Gravity agents can now run [music] on your own computer on a local model with no internet at all. So today I'm building a sky island, a little crew of AI agents
powered by a model that lives on my laptop. Ivy answers client emails. Leo writes openers for your prospects. Nova turns messy coaching notes into a plan, and
Sol builds an SEO dashboard from a search console export. Every one of these runs offline. No API key, no credits. Your client data never leaves your machine. And I'll show you each one working inside Anti-Gravity
step by step so you can copy it today. Here's what Google announced. On the Google Developers blog, the Anti-Gravity team added local model support to the Anti-Gravity SDK. That's the Python kit you use to build your own agents with
the same engine that powers the Anti-Gravity app. The headline setup uses Gemma 4, the 26 billion parameter version running through Google's light RT. Google recommends at least 24 gigabytes of
memory for that one. But here's the bit most people miss. The same post says the SDK works with any OpenAI compatible local server. That means Ollama, LM local server. That means Ollama, LM Studio, or VLLM. So if you're on a
normal laptop like me with 16 gigabytes, you can run a smaller Gemma 4 model through Ollama and use the exact same agent code. Google also showed a hybrid mode. A cloud model plans the job, but it only
sees file names and task descriptions. The local model does the real work on The local model does the real work on your files. In their demo, 97.2% of all the tokens ran locally. So why does this matter for you?
this week to see what people are actually asking. One member asked to get API keys so Claude can run their Gmail automations. Another asked whether their weekly credits roll over because they've been too busy to use them. Someone else
is running DeepSeek locally with their own document search. And one member set up a local vision model to search years of their own video footage. The pattern is clear. People want agents that run all day on their own files without
worrying about keys, limits, or who can see their client data. That's exactly what local Anti-Gravity agents do. You set them up once and they run on your machine whenever you need them. Now, if you want the exact setup from this
video, all four agents, the prompts, and the Sky Island build itself, they're inside the AI Profit boardroom. You get the step-by-step guide for every build plus my faceless video skill and the latest workflows I'm using every day.
And over 3,600 members are in there sharing what they're building, so you're never stuck on your own. The link is in the comments and the description or go the comments and the description or go to ai profit boardroom.com. Right, let's
set it up. Step one is the local model. I've got Ollama installed and I'm using Gemma 4, the E4B version, which runs fine on 16 GB. Step two, I open
Anti-Gravity and start a new conversation in my project folder. Then I simply ask Anti-Gravity to build the agent for me. I tell it to use the Anti-Gravity SDK with the local OpenAI config, point it at Ollama, and pick the
Gemma 4 model. Anti-Gravity creates the virtual environment, installs the Google Anti-Gravity package, and writes a short Python file. The key part is three lines, the model name, the local server address, and the workspace folder the
agent is allowed to work in. That's it. From here, every agent we build uses the same pattern, a folder of inputs, a clear job, and a folder for the results. Here's the exact prompt I gave Anti-Gravity so you can copy it. Build a
Python agent with the Anti-Gravity SDK. Use local OpenAI agent config with the model Gemma 4 E 4 B and the base URL for Ollama on localhost. Set the workspace
to this folder. The agent reads every file in the input folder, does the job I describe, and writes its results to the output folder. Print every tool call so output folder. Print every tool call so I can watch it work. Then run it. That
one prompt is the template for all four agents. I just changed the job description each time. Let's build Ivy, the client inbox desk. This one is for the client inbox desk. This one is for agencies. If you run a marketing, SEO,
or AI automation agency, your inbox is full of client emails that all need a reply. I drop six sample client emails into an inbox folder. Then I ask Anti-Gravity to build an agent that reads every email, ranks them by
priority, writes a one-line summary for each, and drafts a reply in a friendly, clear tone. Anti-Gravity writes the script, runs it, and now watch the local model work. You can see each tool call as it happens. It reads the first email,
then the next, and starts writing files. Here's the result, a triage file with every email ranked, high priority at the top. Email five is from an online store client who's unhappy with their results and thinking about leaving. Ivy spots
it, marks it high priority, and drafts a calm reply that asks exactly where things are falling short, and a drafts folder with a reply for each client ready for me to check and send. All of that ran on my laptop. The client emails
never touched the cloud. Next up is Leo, the prospect opener. This one is for AI automation builders who reach out to new clients. One member said their agent swarm is about to start prospecting, so this one's for them. I give Leo a simple
spreadsheet, eight companies, their niche, and a short note about their website. We 've got a coach, a course creator, a SaaS founder, an agency, and an online store. I ask antigravity to build an agent that writes one personal
first line for each lead on what they actually do. No generic compliments. The local model reads the list and writes an openers file. Each line talks about that specific business. For a course creator with 4,000 students, it picks up on
their busy help inbox. For a SaaS tool with a free trial, it talks about guiding trial users to the right feature. So your first message doesn't sound like everyone else's. Then Nova, the session notes agent. This is for
coaches and consultants. After a call, your notes are usually a mess. Half sentences, ideas, things you promise to send. I add three messy session notes. Nova reads each one and writes a clean plan for every client. Three action
items and one question to open the next session. One client, a newsletter owner, couldn't turn his readers into clients. Nova turned that into three clear steps. A short three email launch sequence, a much shorter sales page, and handing one
much shorter sales page, and handing one inbox task to AI. Plus a question about working. This is where local really matters. Coaching notes are private. With this setup, they stay on your computer and
the agent still does the boring part for you. And finally, Sol, the SEO dashboard. A member asked this week for SEO dashboard ideas. So I built one. I use a sample search console export with 30 rows. Queries, clicks, impressions,
click through rate, and position. Then I ask antigravity for an agent that turns that file into a clean dashboard page. Sol writes a single HTML file. Top queries at the top, then the quick wins. The keywords sitting between position
eight and 20, which are the easiest ones to push onto page one. And a simple bar chart so a client can understand it in 5 seconds. If you run an SEO agency, you can hand this to a client every month and the data never leaves your machine.
Now, two features make these agents safe to leave running. The first is policies. In the SDK, you can block every tool by default, allow reading files, and make the agent ask you before it runs any command. So, Ivy can draft replies, but
she can't delete anything. The second is triggers. You can tell an agent to check something every few minutes. So, Ivy could look at the inbox folder every 10 minutes and draft replies while you're on calls. You can also give these agents
your own tools. Any Python function can become a tool the agent calls, and you can connect MCP servers too. So, a course owner could give the agent a tool that reads their student questions and it drafts answers for the community. An
online store could give it the product spreadsheet and it writes descriptions for every new product. Before you build your own, here are three mistakes to avoid. First, always give the agent its own workspace folder. That keeps it
focused and it can only touch the files you put there. Second, give each agent one clear job with exact file names for the results. Small local models do much better with a tight brief than with a long wish list. And third, give the
model room to think. Google recommends a 64,000 token context for local agents. So, if a long job stops halfway, raise the context size in Ollama before you blame the agent. And if a job is too big for a small local model, use the hybrid
setup from Google's Post. A cloud model plans the steps using only file names and your local crew does the actual work. That's the airship on our island. It drops off the plan and the crew handles everything else. So, which one
should you build first? If you run an agency, start with Ivy. If you build automations for clients, start with Leo. Coaches and consultants start with Nova. And SEO agencies start with Sol because clients love a dashboard they can
actually read. Course and community owners copy Nova's pattern for student questions, and content creators point Leo at your list of brand partners and let it write your outreach openers. Here's my honest take. A small local
model is slower than the big cloud models, and on harder jobs, you'll want to check its work. But for repeat jobs on private files, it's brilliant, and it runs as often as you like. If you want all four agents, the setup guide and the
exact prompts I used, grab them inside the AI Profit Boardroom. You get the full anti-gravity local setup, my faceless video skill, and new workflows added every day. The link is in the comments and the description, or go to
aiprofitboardroom.com. Thanks for watching, and I'll see you in Thanks for watching, and I'll see you in the next one.
