Install & Start Your First Agent
Last reviewed
BeginnerWhat you'll learn
~20 min- Choose a desktop agent app or a terminal CLI, on Windows, macOS or Linux
- Install one tool, sign in, and open it on a practice folder called ai-class
- Switch to plan mode inside the session, answer the agent's questions, and approve its plan
- Open the page your agent built in your own browser, and ask for one change
In the next 20 minutes you install a real workspace agent on your computer, open it on a practice folder, and watch it build a small web page that you then open in your own browser. No coding needed. You type a few short lines at most; everything after that is plain English.
The last lesson described three doors to the same agent. This lesson walks you through one of the two local doors, today, so every lesson from Module 2 onward runs in your own agent on your own machine.
If you cannot install software on this computer, use a browser or cloud builder instead: Choosing Your Cloud Sandbox shows the options. The plan, build and check steps below work the same way there.
Step 1: Pick one tool and choose your door
Each of the four makers ships the same agent as a desktop app and as a terminal CLI. Both work on a folder on your computer, both plan with you before they build, and you can switch between them later without losing anything, because your work is just files in a folder. Pick the door that feels comfortable today; neither is the beginner version of the other.
| Tool | Desktop app | Terminal CLI | Free way in (as of November 2026) |
|---|---|---|---|
| Antigravity (Google) | Antigravity app | agy | Yes: free tier with a Google account, under weekly limits (details) |
| GitHub Copilot | GitHub Copilot app | copilot | Yes: Copilot Free covers the app and the CLI (plans) |
| Codex (OpenAI) | ChatGPT desktop app, Codex or Work | codex | Yes, in the desktop app: Free and Go include GPT-6 Luna there (pricing) |
| Claude Code (Anthropic) | Claude desktop app, Code tab | claude | No: Pro ($20/month) or a higher plan (pricing) |
If you already pay for one of these, use that one. If not, Antigravity and GitHub Copilot cost nothing to start, through either door.
Everything in this course uses made-up example data. The agent sends what it reads to the company that runs it, so keep personal, confidential or work files out of your practice folder.
Door A: a desktop agent app (no terminal)
- Install one app from its maker’s page and sign in:
- Antigravity
- GitHub Copilot app
- ChatGPT desktop app (use Codex or Work)
- Claude desktop app (use the Code tab)
- In Finder (Mac) or File Explorer (Windows), open your home folder (the one with your name) and make a new folder called
ai-class. - Start a new session in the app, choose the
ai-classfolder when it asks for a folder, project or workspace, and check that the app names it before you type.
The app may ask before it changes a file or runs a command, with a button to allow it. Inside this practice folder it is fine to allow them. Now skip to Step 2.
Door B: a terminal CLI
Open a terminal, the window where you type commands:
- Windows: press Win + X and choose Terminal (on some computers it says Windows PowerShell). The line waiting for you starts with
PS. - macOS: press Cmd + Space, type Terminal, and press Enter.
- Linux: press Ctrl + Alt + T.
Install one tool. Copy the line for your tool, paste it at the prompt, and press Enter. Each line fetches the maker’s own installer and runs it.
Antigravity CLI:
curl -fsSL https://antigravity.google/cli/install.sh | bashGitHub Copilot CLI:
curl -fsSL https://gh.io/copilot-install | bashCodex CLI:
curl -fsSL https://chatgpt.com/codex/install.sh | shClaude Code:
curl -fsSL https://claude.ai/install.sh | bashIn Windows PowerShell:
Antigravity CLI:
irm https://antigravity.google/cli/install.ps1 | iexGitHub Copilot CLI needs PowerShell 7, a newer PowerShell than the one Windows comes with. Install it first:
winget install Microsoft.PowerShellThen open PowerShell 7 from the Start menu, use it for the rest of this lesson, and install Copilot CLI:
winget install GitHub.CopilotCodex CLI:
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"Claude Code:
irm https://claude.ai/install.ps1 | iexAlready using Ubuntu (WSL)? You can use it instead and follow the Linux tab.
Linux, or Ubuntu (WSL) on Windows:
Antigravity CLI:
curl -fsSL https://antigravity.google/cli/install.sh | bashGitHub Copilot CLI:
curl -fsSL https://gh.io/copilot-install | bashCodex CLI:
curl -fsSL https://chatgpt.com/codex/install.sh | shClaude Code:
curl -fsSL https://claude.ai/install.sh | bashWhen it finishes, close the terminal and open a new one, so it finds the tool you just installed.
Make the practice folder and go into it. Type these two lines one at a time, pressing Enter after each (they work the same on Windows, macOS and Linux):
mkdir ai-classcd ai-classStart your tool. Type only the one line for your tool. The part starting with -- is a flag, a setting for this one start, and these flags turn the approval prompts off: the agent then creates, changes and deletes files and runs commands without asking you first, and depending on the tool it can reach other files your account can open, so start it only inside ai-class and keep anything you cannot replace somewhere else.
Antigravity CLI:
agy --dangerously-skip-permissionsGitHub Copilot CLI:
copilot --allow-allCodex CLI:
codex --dangerously-bypass-approvals-and-sandboxClaude Code:
claude --dangerously-skip-permissionsIf you would rather the tool asked you more often, start it with just its name (agy, copilot, codex or claude); Module 6: Your Agent In Depth explains what each one does then. The first start may show a few screens before you can type: signing in through your browser, a colour theme, its terms, or a question about trusting this folder. If it asks about a folder, check that it names ai-class, accept, and wait for the box where you type your message.
Step 2: Run your first real task
From here on you type into the agent, in plain English, not into the terminal.
Switch to plan mode. In a terminal CLI, press Shift+Tab until the bottom of the screen says plan, or type /plan. In a desktop app, pick plan from the mode menu beside or under the message box (if you cannot find it, ask the app how to switch to plan mode). In plan mode the agent reads, asks its questions, and shows you a plan before it builds.
Ask for a small page. Say it your own way; something like:
Make a small welcome.html page in this folder: a clean quick-reference card for this AI course, with space for my name and three tips. Tell me how to open it in my browser.Answer its question. It may ask about colours, which tips to include, or your name. Reply briefly, the way you would answer a colleague:
Dark blue and white. My name is Sam. Pick three beginner tips yourself.Approve the plan. It comes back with a short plan, usually one file. Read it. If the tool offers a choice to go ahead, pick it. Otherwise say “That looks right, go ahead.” (In Codex, press Shift+Tab first to leave plan mode, because in plan mode Codex treats a request to build as a request to plan it.)
Open what your machine just built. Follow the agent’s directions, or open your ai-class folder in Finder or File Explorer and double-click welcome.html. It opens in your browser. That page was written to your disk by the agent you installed ten minutes ago.
Ask for one change. Back in the agent:
Make the tips into a numbered list and add today's date at the top.Refresh the page in your browser and check the change. That is the build loop this course teaches: plan, approve, build, check, keep (keep comes in Module 7).
To leave a terminal CLI, type /exit (in Antigravity, press Ctrl+D with nothing typed), or close the window. In a desktop app, close the session. Your ai-class folder and the page stay where they are.
Module 6: Your Agent In Depth: Desktop App & CLI has a lesson for each of the four ecosystems: plans and models, every mode and what it allows, switching between the desktop app and the CLI, and troubleshooting.
Common issues on your first start
Use the Symptom → Evidence → Request pattern: describe what you see, paste the error, then ask for a fix.
Key takeaways
- Two local doors, one agent: a desktop app or a terminal CLI, on Windows, macOS or Linux; both work on the same folder and you can switch later
- The terminal part is a few lines: install, make a folder, go into it, start the tool; everything after that is plain English
- Approval prompts off means no questions asked, so the practice folder holds only practice files
- Plan first, then build, then check: you switched to plan mode, answered a question, approved, and opened the result yourself
- Your agent is now your classroom: from Module 2 on, every hands-on step runs in it
Your Agent Tutor
Below this lesson is the Agent Tutor card, and from Module 2 onward a copy sits at the top of each lesson. Copy its one prompt into your agent, in plan mode, inside that lesson’s folder. Your agent then reads the tutor page and the lesson itself, and coaches you through the steps one at a time on your own machine.
You started your terminal CLI with approval prompts off. What does that change?
Your agent built welcome.html. What is the best way to check its work?