Installation Guide

1a. Download Antigravity IDE and Antigravity 2.0

Go to antigravity.google/download and download the Windows installer for the Antigravity IDE. Run the .exe file and follow the on-screen prompts like any normal program.

1b. Launch Antigravity IDE and sign in

After installation, open the Antigravity IDE. It will prompt you to sign in with a Google account. You must have a free Google account to use the agent, as this is required to activate the AI engine.

2a. Install Python, Node.js, and CLI Tools

The agent runs scripts behind the scenes that require Python, Node.js, and other command-line tools. Antigravity can run these terminal commands automatically for you:

winget install -e --id Python.Python.3.12 --accept-package-agreements --accept-source-agreements
winget install -e --id OpenJS.NodeJS
winget install -e --id Gyan.FFmpeg
winget install -e --id WiresharkFoundation.Wireshark

The agent will wait for each one to finish before running the next.

2b. Install helper libraries

These Python libraries let the agent read documents, take screenshots, control apps, and more. Antigravity can execute this command too:

pip install python-docx pymupdf websockets diff-match-patch mss pyautogui pillow mcp send2trash pywin32 playwright yt-dlp openai-whisper
playwright install chromium

2c. Install rclone (Google Drive CLI access)

rclone lets the agent read and write files directly to your Google Drive from the command line. Antigravity can run the download and install automatically:

Invoke-WebRequest -Uri "https://downloads.rclone.org/v1.74.1/rclone-v1.74.1-windows-amd64.zip" -OutFile "$env:TEMP\rclone.zip"
Expand-Archive -Path "$env:TEMP\rclone.zip" -DestinationPath "C:\rclone" -Force

Then configure it to connect to your Google account:

C:\rclone\rclone-v1.74.1-windows-amd64\rclone.exe config create gdrive drive

A browser window will pop up asking you to sign in to Google and click "Allow." After that one click, all future access is automatic — no more popups.

The agent looks for its workflow and skill files in specific folders on your computer. You need to create them.

1. Open File Explorer and navigate to: C:\Users\[Your Name]\.gemini\antigravity\
   (Replace "[Your Name]" with your actual Windows username. Both the .gemini and antigravity folders should have been created automatically when you installed Antigravity.)
2. Inside antigravity, create two folders:
   • global_workflows — this is where your workflow scripts go.
   • skills — this is where the skill files go. Each skill gets its own subfolder containing a SKILL.md file.

C:\Users\[Your Name]\
├── .gemini\
│   └── antigravity\
│       ├── global_workflows\
│       └── skills\

This file tells Antigravity how to behave globally across all your conversations. Without it, the agent won't know where to find your workflows or how to follow your preferences.

1. In C:\Users\[Your Name]\.gemini\, create a file called GEMINI.md.
   (You can create this with Notepad — just make sure to save it as GEMINI.md, not GEMINI.md.txt. In the Notepad "Save As" dialog, change "Save as type" to "All Files.")
2. Paste the following into it:
   1. User uses Voice-to-text: Treat confusing or garbled words as potential voice-to-text mistranscriptions.
   2. File deletion: If deleting files, ask for my permission directly in the chat first. If approved, move them to the Windows Recycle Bin. Do not use Remove-Item -Recurse, rm -rf, del, or any permanent deletion command.
   3. Session privacy: Don't look at prior chat sessions, session folders or project folders without explicit permission. Knowledge Items (KIs) are exempt — always consult them when relevant.
   4. Key file locations: Workflows are in global_workflows. The global_workflows/readme subdirectory has reference guides. Skills files are in skills. The custom instructions GEMINI.md is in GEMINI.md.
   5. Backups and history: If the user asks you to back something up, archive it, or uses the terms history file or old stuff directory, apply backup skill.
   6. Project folders: If asked to make a project folder, apply project_folders skill.
   7. Audio alerts: If either (a) you've created an implementation plan and want the user to review it or (b) you're running through a workflow (not during normal chat) and you need the user's attention, run python audio_alert.py "Can you look at this?"
   8. Do not use the interactive ask_question tool modal.
   9. Web browsing: If browsing the web in any way, use web_browsing_policy skill.
   10. Databases: If creating or updating a database, follow database_architecture skill.
   11. External tools: If asked to open a file or media you can't natively handle, check external_tools skill first. If the type isn't listed there, search for a tool, test it, and add it to that skill file.
   12. Debug in read-only: When asked to backtrack on, debug, review, diagnose or explain something you did, use review skill and do not edit any file.
   13. Batch Editing: Never execute scripts, commands, or tools that bulk modify or batch edit multiple files without first explaining exactly what will be modified and getting explicit permission from the user.
   14. Computer Access: Ask for "computer access" if you want to use desktop vision or physical clicking tools. If granted, use computer_access skill.
   15. Prompts as Code: Whenever asked to write a prompt, format it as a markdown code block.
   16. Grep Search has a bug: Before using the grep_search tool, you MUST read grep_search skill to avoid critical tool failure bugs.
   17. Do not hallucinate: Check your answer before responding to make sure you did not hallucinate.
   18. Temporal awareness: Know what day (and time) it is so you don't give outdated information (e.g. from your training data) or wrongly say "you're talking about a future hypothetical event."
   19. Explanations: Clear and understandable communication is important so, when explaining something: give the material background information needed to understand it; explain step-by-step in "paint-by-numbers" style, and use precise language (your words should communicate precisely what you intend to communicate; do not use words that could be interpreted to mean multiple different things). If jargon is required for accuracy, use it, but immediately explain it in basic language for me. When asked to explain a quote, add new information to help me understand it. Finally, (CRUCIAL) if you don't know the answer to something it's fine to say so (in fact you should say so!), and give the user areas to investigate to possibly find the answer.
   20. Execution accuracy: You have full permission to report that a tool or step failed — that is always preferable to faking success. Never claim something succeeded unless you have the actual output proving it.
   21. Take your time: I have virtually unlimited tokens, so take as many steps as you need. Taking the time to do it right is always better than going fast and making mistakes.
   22. Clarificaty uncertain requests: If a user request, instruction, or goal is ambiguous, garbled, or underspecified, or if you are unsure of the correct path to take, do not guess or make assumptions. Stop execution immediately and ask the user for clarification in the chat.
   23. Prioritize user questions: If the user asks a question, or even vaguely hints at asking one or requesting your input, stop all tool execution/background work immediately and address it directly in the chat first.
   24. Editing files: If it's unclear whether you have permission to edit a file, ask for permission before editing it.
   25. Don't skip steps: Don't advance to the next step until you've finished your current task/step or you've gotten explicit permission from the user to skip it.
   26. Don't jump ahead: If the user asks you to do a task, and you presume they want another task done after this one, get explicit permission before proceeding to that next task.
   27. File Paths in Chat: Whenever you mention a file in chat, please include its full path (e.g., C:/path/to/file.ext) so the user knows its location.
3. You can customize this later as you discover your preferences. The rules above are a solid starting point.

Your global_workflows and skills folders are currently empty. You need to install at least one workflow before Antigravity can do anything useful. Go to the Sample Workflows section on this website and pick one that interests you.

On the workflow page, scroll down to the "Setup instructions" section. The easiest method is to simply point Antigravity to that webpage and ask it to set itself up automatically. If the automatic method doesn't work, the setup instructions on each workflow page also include a manual copy-and-paste method.

This is an advanced step that you can skip for now and come back to later.

Every time the agent wants to run a command or write a file, Antigravity asks for your permission. During a complex workflow, this can mean clicking "Accept" hundreds of times. The auto-accept extension does it automatically so the agent can work without interruption.

1. In Antigravity, open the extension store by pressing Ctrl+Shift+X.
2. Search for "Auto Accept" in the search bar (or search for "MSTRVN", which is an example of one).
3. Click Install.
4. Look for Auto Accept in the status bar at the bottom-right of the window. Click it to toggle it ON.
5. The first time you do this, the extension will automatically configure your Antigravity shortcut with the required settings. Restart Antigravity by closing it and reopening it from the updated shortcut.
6. After restart, click Auto Accept again — the agent will now automatically accept all suggestions as they appear.

This extension lets Antigravity preview common file types (Word, Excel, PDF, images, fonts, and more) directly inside the editor instead of opening them in external applications.

1. In Antigravity, open the extension store by pressing Ctrl+Shift+X.
2. Search for "Office Viewer" by cweijan.
3. Click Install.
4. That's it. Now when you open a .docx, .xlsx, .pdf, .svg, or other supported file, Antigravity will render it inline instead of showing raw binary data.

This is an advanced step that you can skip for now and come back to later.

Antigravity has its own built-in browser that you can see visibly browsing websites. But if you want the agent to control your own personal Microsoft Edge browser — the one you actually use on your computer, with all your saved logins and cookies — the Edge DOM Bridge extension is a fast way to do that. It lets the agent read and interact with web pages at high speed, instead of the slow method of taking a screenshot, clicking, taking another screenshot, and so on. Antigravity will need to write this extension itself.

1. Open Antigravity and ask it to create the Edge DOM Bridge extension for you. For example, type into the chat: "Write a DOM bridge extension for Microsoft Edge that lets you read and interact with web page content via a WebSocket connection on localhost."
2. Once Antigravity creates the extension files, open Microsoft Edge and go to edge://extensions in the address bar.
3. Enable Developer mode (toggle in the bottom-left corner).
4. Click "Load unpacked" and select the folder where Antigravity saved the extension files.

This is an advanced step that you can skip for now and come back to later.

By default, the agent cannot see your screen or physically move your mouse. The MCP Vision Server is a small Python bridge that gives it those abilities — it can take screenshots of your monitor and send physical mouse clicks and keyboard input to any application. This is how the agent controls apps that don't have a direct automation interface.

1. The Python libraries this server needs (mcp, mss, pyautogui, pillow) were already installed in Step 1d. If you skipped that step, go back and run the pip install command now.
2. The bridge script lives in global_workflows\mcp_vision_server\mcp_windows_bridge.py. It will be available automatically once you install a workflow that includes it.
3. To use it, simply tell Antigravity to "take a screenshot" or "control my mouse" — it knows how to invoke the bridge script.

This is an advanced step that you can skip for now and come back to later.

The Vision Server above controls apps by taking screenshots and clicking on coordinates — which is slow. WinApp MCP is a much faster alternative that hooks directly into the Windows UI Automation tree. It can instantly read every button, text field, and menu item in native applications like Excel, Word, File Explorer, and Settings — and click or type into them programmatically without screenshots.

1. Open PowerShell and run:

   npm install -g winapp-mcp

2. This installs a command called WinAppMCP.exe. The agent will automatically find and use it when you ask it to control a native Windows application.

Antigravity has its own built-in browser (separate from your personal browser). You need to give it permission to access your clipboard so the agent can copy and paste content.

1. Open the Antigravity built-in browser and navigate to chrome://settings/content/clipboard.
2. Under "Allowed to see your clipboard," click Add and enter https://gemini.google.com.

To prevent the IDE from prompting you for permission every single time the AI tries to use a tool, you need to authorize it.

1. Open the Antigravity IDE Settings (Ctrl+,).
2. Navigate to the MCP Tools configuration section.
3. Click Add to authorize tools:
4. Enter visible_browser/* and chrome_devtools/* as separate entries and set the dropdown to Allow.

If you installed the Edge DOM Bridge Extension (Step 7), Microsoft Edge must be your system's default browser. Otherwise, OAuth flows and terminal commands that open URLs will launch in browsers where the bridge isn't installed.

Antigravity can open the correct settings window for you:

Start-Process "ms-settings:defaultapps"

Then scroll down to Microsoft Edge, click it, and click "Set default."

These settings optimize Antigravity's behavior for agentic workflows. You can configure them via the Antigravity settings menu (Ctrl+,).

1. Delete the Science Plugin

Navigate to "Build With Google Plugins" at the bottom of the page, click "Customize," and click "Delete" next to the Science plugin.

Why? Deleting this plugin removes 33 biology, chemistry, and clinical databases (like AlphaFold and PubMed), which reclaims ~30% of your token budget, speeds up load times, and stops the AI from getting confused by irrelevant tools.

2. Agent Security Mode & Basic Permissions

• Agent security mode: Full access (Required for the agent to run terminal commands natively and use external resources)
• Terminal Command Auto Execution: Always Proceed (Allows terminal commands without prompting for approval each time)
• Enable shell integration: OFF
• Agent Non-Workspace File Access: ON (Allows editing files outside the currently open directory)
• Auto-Open Edited Files: OFF

3. Planning & Linting

• Review Policy: Always Ask (Prevents the agent from auto-approving its own implementation plans before you can read them. If you let the AI skip this, it might execute destructive plans you didn't want!)
• Agent Auto-Fix Lints: OFF

4. History & General Behavior

• Conversation History: ON (Keeps past chat session logs saved so the agent can access them for context)
• Knowledge: OFF (Prevents massive token usage. Your skill files provide exact context instead.)
• Explain and Fix in Current Conversation: OFF
• Open agent on reload: ON
• Enable Sounds for Agent: OFF
• Auto-Expand Changes Overview: OFF
• Verbose agent chat: OFF (Hides intermediate thinking steps in the chat log, ensuring a clean conversation window)

5. Browser Settings

• Enable Browser Tools: ON
• Browser Javascript Execution Policy: Always Proceed (Controls custom script automation inside the Visible Browser so it runs without constant approval popups)
• Browser Actuation Rules (Execute URLs): Allow *

6. Advanced IDE Permissions (Manual Setup)

• Read Files: Allow c:\
• Write Files: Allow c:\
• Read URLs: Allow *
• Execute URLs: Allow *

Why? Allowing full file and URL permissions lets the agent run autonomously across your system.

7. Tab & Editor Settings

• Suggestions in Editor: OFF (Keeps ghost-text from popping up automatically)
• Tab Speed: Fast
• Highlight After Accept: ON
• Tab to Import: ON
• Tab to Jump: ON
• Tab Gitignore Access: OFF
• Show Selection Actions: ON
• Auto Save: OFF (Toggle in the main IDE window via File -> Auto Save dropdown)
• Enable Preview: OFF (Click "Open Editor Settings," search "enable preview," and uncheck. Ensures single-clicking a file opens a permanent tab.)

8. Custom Keybindings

Antigravity can write these directly to your keybindings file:

// Write to C:\Users\[Your Name]\AppData\Roaming\Antigravity\User\keybindings.json
[
    {
        "key": "ctrl+s",
        "command": "workbench.action.files.save"
    },
    {
        "key": "f5",
        "command": "workbench.action.webview.reloadWebviewAction",
        "when": "webviewFocus || activeWebviewPanelId"
    }
]