Terminal AI Agents — A Complete Beginner's Guide

Written so that you can follow along without being good with computers.
Based on Windows 11. Work through it in order, start to finish.
Estimated time: 30 minutes to install + 20 minutes for your first session

How to read this document

• Want to start right now? Follow The short version — the fastest path through below. Open the relevant section only when you get stuck.
• First time? Start at Part 1 and read in order. Do not skip.
• Already used a terminal? Skip Part 2 and start at Part 3.
• Text in grey boxes is a command you type into the computer. Copy and paste it as is.
• Every step has a "Check that it worked" item. Confirm it before moving on.
• Appendix E holds optional settings you only need if something breaks. Skip it on a first pass.

The short version — the fastest path through

If you have no time to read the whole thing, just follow this sequence. The section number at the end of each step is where the full explanation and troubleshooting live.

1. Open a terminal

1. Press the Windows key on your keyboard (the one with the window icon, usually next to the left Ctrl).
2. Type terminal into the search box that appears.
3. When the app named Terminal shows up, press Enter.

A black (or blue) window opens. If you see text like this, it worked.

PS C:\Users\Jane>

→ In full: section 2.1

2. Install Node.js

Type this into the terminal and press Enter.

winget install --id OpenJS.NodeJS.LTS -e

Check — close the terminal, open a new one, and type this.

node --version

Output like this means it worked. The numbers will differ.

v24.9.0

→ In full: section 3.2

3. Install Git

Type this into the terminal and press Enter.

winget install --id Git.Git -e

Check — open a new terminal and type this.

git --version

git version 2.52.0.windows.1

Output like that means you are set. If nothing comes back, open a fresh terminal and install again.

→ In full: section 3.3

4. Install an AI agent

Install only the one you subscribe to.

a. Claude Code (Claude subscribers)

Type this into the terminal.

irm https://claude.ai/install.ps1 | iex

If that fails, type this instead.

npm install -g @anthropic-ai/claude-code

Check — close the terminal completely, open a new one, and type this.

claude --version

b. Codex CLI (ChatGPT subscribers)

Type this into the terminal.

irm https://chatgpt.com/codex/install.ps1 | iex

If that fails,

npm install -g @openai/codex

If that fails too,

winget install --id OpenAI.Codex -e

Check — close the terminal completely, open a new one, and type this.

codex --version

→ In full: section 4.2 (Claude Code) and section 4.3 (Codex CLI)

5. Log in and start

First make a working folder and move into it. Do not launch it on your Desktop or in C:\Users\yourname.

mkdir C:\ai-practice

cd C:\ai-practice

From there, type whichever one you installed and press Enter.

claude

codex

The first time, a login prompt appears. Sign in with your subscription account in the browser and come back to the terminal. After that it starts straight away.

→ In full: sections 5.1 to 5.3

6. Run several agents at once with Orca (optional)

Orca is an app for running several AI coding agents at the same time. If you finished step 4, you are already set up.

a. Install

1. Go to https://www.onorca.dev in a web browser.
2. Press the Download button.
3. Download the Windows installer (orca-windows-setup.exe).
4. Double-click the downloaded file.
5. When the installer window opens, keep the defaults and continue.

b. Use it

1. Next to Projects on the left, press the folder icon with a + (Add Project) and pick the folder to work in. That folder must be a Git repository — if it is not, open a terminal there and run git init once.
2. Press the + above the project name (New workspace) to create a workspace.
3. Pick the agent you want from the dropdown on the run screen and give it instructions as usual.
4. Create a second and third workspace, and that is parallel work.

Switch the approval mode to Manual at first. Orca ships with permission bypass on by default.

→ In full: Orca beginner's guide

7. The five essential commands

Inside the chat window these five are effectively all you use. Know them and you will never be stuck.

→ In full, with usage limits: section 5.0

8. Extend it with MCP servers and plugins (optional)

Korean HWP documents, statutes, national statistics — things the AI cannot do out of the box get added as extensions. You do not need to find the install command; give it the repository address and ask.

Install https://github.com/chrisryugj/kordoc globally so I can use it.

When it finishes, close the terminal completely and open a new one.

→ In full: Part 7

Part 1. Why run AI in a terminal

1.1 How this differs from the AI chatbot you already use

When you use ChatGPT or Claude in a web browser, the flow looks like this.

You: "Tell me how to clean up this Excel file"
AI:  (explains the method in words)
You: (read it -> open Excel -> do the work by hand)

The AI only talks. You do the actual work. You upload files one at a time, then copy the answer back out by hand.

A terminal AI agent works differently.

You: "Clean up the Excel files in this folder"
AI:  (opens the folder -> reads the files -> edits them -> reports back)
You: (just review it)

There is one core difference: the AI can read and write files on your computer directly.

1.2 What you can actually do with it

Things you can do without being a programmer, grouped by area.

Documents and reference material

• Turn 100 scattered PDFs into a table of titles and summaries
• Rename files in bulk to a consistent convention
• Pull one specific section out of many Word files and merge them into one

Data work

• Find and fix errors in Excel/CSV files
• Merge several tables into one and compute statistics
• Plot data as a chart and save it as an image file

Research and academic work

• Rewrite a draft's sentences in an academic register
• Normalise references to a journal's required style
• Write and run a script that processes experimental data

Everyday office work

• Fill in a recurring report template automatically
• Gather material from the web and organise it
• Draft many emails at once

Programming (if you are interested)

• Build a program from scratch
• Understand and modify code someone else wrote

1.3 Answers to the usual worries

"Do I need to know how to program?"

No. You say "do this for me" in plain English. You do need to know how to open a terminal and a handful of commands, which Part 2 covers. Thirty minutes is enough.

"What if it wrecks my files?"

An AI agent asks "may I make this change?" every single time before it modifies a file. Nothing runs without your approval. On top of that, if you follow the safety rules in Part 8 (backups, a dedicated folder), the risk is close to zero.

"Will my files leak?"

File contents the AI reads are sent to a server so it can produce an answer. This is the same as uploading a file to a web chatbot. Do not put personal data, unpublished research data, or company confidential material in your working folder. Details in Part 8.

"Does it cost money?"

A paid subscription is generally required. If you already pay for ChatGPT Plus or Claude Pro, you can use it through that subscription at no extra cost. Google's Antigravity CLI gives you a free allowance when you sign in with a Google account (check the official documentation for the exact limits). Section 4.1 lays this out.

"Do I need Linux? Do I need WSL?"

No. Use it on Windows as is. Some tools did not support Windows in the past and required WSL (a Linux environment), and a lot of material from that era is still circulating. All three tools now ship a Windows-native installer. See Appendix D for how to decide.

"The black screen scares me."

That is the most common reaction. In practice you use about five commands; everything else is a plain-English conversation. Think of the black screen as just a window for talking to the AI.

Part 2. Terminal basics — 30 minutes is enough

You are not trying to become a programmer. You learn just enough to give the AI work.

2.1 What a terminal is

A terminal is a window where you give the computer instructions as text.

Normally you open a folder by double-clicking it with the mouse. In a terminal you type cd foldername instead. The result is the same. Only the method differs.

The reason this matters is that AI agents only work this way. There is no screen to click.

Open a terminal

1. Press the Windows key on your keyboard (the one with the window icon, usually next to the left Ctrl).
2. Type terminal into the search box that appears.
3. When the app named Terminal shows up, press Enter.

A black (or blue) window opens with text like this.

PS C:\Users\Jane>

This is normal. Here is what it means.

• PS = short for PowerShell. The name of the program taking your commands.
• C:\Users\Jane = where you currently are (which folder). This is the important part.
• > = the signal that says "type a command here". You type after it.

Remember this: a terminal always has a "current location". It is exactly like having a particular folder open in File Explorer. Every command runs relative to that location.

2.2 The shape of a command

Most commands look like this.

command target -option

Here is an example.

cd Documents

• cd = the command (means "move into this folder", short for change directory)
• Documents = the target (the folder to move into)

After typing a command you must press Enter for it to run. Nothing happens until you do, so if you make a typo, do not panic — just delete it.

2.3 Understanding folders and paths

Folders on a computer nest inside each other like a chest of drawers.

C:\                          <- the C: drive (outermost)
+-- Users\                   <- the users folder
    +-- Jane\                <- your folder
        +-- Documents\
        +-- Downloads\
        +-- Desktop\

Writing that out on one line gives you a path.

C:\Users\Jane\Documents

The backslash (\) is the separator between drawers. On some non-English keyboards it is displayed as a currency symbol (for example the won sign on Korean keyboards), but it is the same character. Do not worry about it.

How to avoid typing paths by hand (important)

Typing paths by hand invites typos. There are much easier ways.

Method 1 — copy it from File Explorer
1. Open the folder you want in File Explorer.
2. Click once on the address bar at the top. The path turns into text.
3. Copy it with Ctrl + C.
4. In the terminal, type cd (with a space after cd), then click the right mouse button once to paste.

Method 2 — open a terminal straight from Explorer (easiest)
1. Open the folder you want in File Explorer.
2. Right-click on empty space inside the folder.
3. Choose "Open in Terminal".

The terminal opens already at that folder. This is the method you will use most.

Careful: if the path contains spaces, wrap it in double quotes.
Wrong: cd C:\My Documents\Research Data
Right: cd "C:\My Documents\Research Data"

2.4 The five commands you must know

This is all you need.

Run these in PowerShell.
The commands below work in PowerShell, which is what the Terminal app you opened in section 2.1 uses by default. In the old Command Prompt (cmd) there is no ls or pwd, so they fail. If your prompt starts with PS C:\...> you are in PowerShell; if it starts with just C:\...> you are in Command Prompt — close it and open the Terminal app instead.

1. cd — move into a folder

cd Documents

Moves into the Documents folder.

cd ..

Two dots means go up one level (outwards).

cd \

Goes to the very top of the C: drive.

2. ls — see what is in the current folder

ls

Lists the files and folders where you are. The same as looking inside a folder in Explorer.

3. pwd — check where you are

pwd

Shows the full path of your current location. Use it when you get lost.

4. mkdir — make a new folder

mkdir myresearch

Creates a folder called myresearch where you are.

5. cls — clear the screen

cls

Tidies up a cluttered screen. It clears the display only — no files are deleted.

2.5 Handy things to know

Do learn the Tab key. Type cd Doc, press Tab, and it completes to cd Documents\ by itself. It prevents typos.

2.6 Practise right now

Open a terminal and type the following in order. Press Enter after each line.

pwd

-> shows where you are.

ls

-> shows a file list.

cd Documents

-> moves into the documents folder.

pwd

-> you can see the location has changed.

mkdir ai-practice

-> creates a new folder.

ls

-> the ai-practice folder you just made appears in the list.

cd ..

-> back out again.

If you got this far, terminal basics are done. That really is all of it.

When something does not work

No error will break your computer. Try things freely.

Part 3. Preparation — three supporting programs

Before installing an AI agent, install the programs it runs on.

3.0 What you are installing and why

All of these are free, safe, official programs.

The execution-policy and encoding settings are not done here.
Earlier versions of this guide told you to set them up in advance. Today's installers almost always work without them, and there is no reason to change settings you do not need, so they now live in Appendix E. If an install or a run fails with running scripts is disabled, or non-English text shows as ???, go to Appendix E at that point.

Why Node.js is only "recommended"
AI agents used to be installable only through npm (the installer that ships with Node.js). Today Claude Code, Antigravity CLI and Codex CLI all provide dedicated installers that work without Node.js. It is still worth having: it gives you a fallback route when an install goes wrong, and a good share of the tools the AI writes for you will use it. If you are short on time, skip section 3.2 and go to Part 4.

3.1 Install Windows Terminal

Windows 11 nearly always has it already. If the terminal opened fine in section 2.1, skip this step.

If not, type this into a terminal (or Command Prompt).

winget install --id Microsoft.WindowsTerminal -e

3.2 Install Node.js (recommended)

Node.js is the underlying engine that runs and installs programs. It is used when you install an AI agent the npm way, and often needed to run the tools the AI builds for you.

Method A — by command (recommended, fast)

Type this into the terminal and press Enter.

winget install --id OpenJS.NodeJS.LTS -e

It takes a few minutes. When you see Successfully installed, it is done.

Method B — download from the website

Use this if the command does not work.

1. Go to https://nodejs.org in a browser.
2. Click the green button marked "LTS". (LTS = the long-term support stable release. Do not pick Current.)
3. Double-click the downloaded file (node-v24...msi).
4. When the installer opens, change nothing and keep pressing "Next".
5. If a checkbox reading Automatically install the necessary tools... appears, leave it unchecked. It is unnecessary and slow.
6. "Install" -> then "Finish".

Check that it worked (do not skip)

Close the terminal window completely and open a new one. (Skip this and it will look like the install failed.)

In the new terminal, type:

node --version

Success looks like this.

v24.9.0

The numbers can differ. Any version number starting with v is fine. (As of September 2026 the LTS line is 24.x, and 22.x is still supported.)

npm --version

Check this too. A number like 10.9.2 means you are fine.

What is node and what is npm?
node is the engine; npm is the tool that installs programs. Installing Node.js gives you both.

When it does not work

If you see 'node' is not recognized...:

1. Confirm you closed and reopened the terminal. (This is the most common cause.)
2. If that fails, reboot and check again.
3. Still failing? See section 9.1.

3.3 Install Git

Git is a tool programmers use to track file history, and AI agents use it internally. You do not need to learn it — just have it installed.

On Windows there is a second reason. Installing Git also installs Git Bash, which Claude Code detects automatically and uses as its command shell. The Linux-style commands that fill the internet then work as written, and errors drop noticeably. It is not mandatory, but you are better off with it.

winget install --id Git.Git -e

To get it from the website, download from https://git-scm.com/download/win and install. The installer asks a lot of questions; leaving every default alone and pressing Next is fine.

Check

Open a new terminal and type:

git --version

git version 2.52.0.windows.1

Output like that means you are set.

3.4 Part 3 final check

Open a new terminal and type each of these, confirming that a version number comes back.

node --version

npm --version

git --version

If all three return version numbers, you are ready. If any one fails, see Part 9.

Part 4. Installing an AI agent

4.1 Which one to choose

Pick one of three. If this is your first time, install one and get comfortable before adding more.

Recommendations

• I want to try it without paying -> Antigravity CLI (section 4.4)
• I already pay for ChatGPT Plus -> Codex CLI (section 4.3)
• I already pay for Claude Pro, or documents and research are my main use -> Claude Code (section 4.2)

Installing all three causes no conflicts. Still, start with one.

If you came here looking for Gemini CLI
Google's gemini command (Gemini CLI) stopped serving individual users on 18 June 2026. Its successor is Antigravity CLI (agy), covered in section 4.4. Older material online still tells you to run npm install -g @google/gemini-cli; install it that way today and you will be blocked at the login step. Section 4.4 has the details.

4.2 Install Claude Code

Method A — dedicated installer (recommended)

Type this into the terminal.

irm https://claude.ai/install.ps1 | iex

It takes one to three minutes. No Node.js needed, and it keeps itself updated. Use this if you are starting out.

irm ... | iex means "download an install script from the internet and run it immediately." Check that the address is claude.ai before you run it. Never use this pattern with an address you do not trust.

Method B — via npm (if you already have Node.js)

npm install -g @anthropic-ai/claude-code

Lines of text scrolling past is normal. A few yellow WARN messages can be ignored. A red ERR! means you should look at Part 9. Requires Node.js 22 or later.

-g means "install for the whole machine." You need it so the command works from any folder.
Do not use sudo npm install -g. It creates permission problems.

Do only one of the two. If you already installed via npm, run claude install to migrate to the dedicated installer.

Check that it worked

Close the terminal completely, open a new one, then type:

claude --version

A version number followed by (Claude Code) means success.

Log in

Make a folder to work in and run it there.

mkdir C:\ai-work

cd C:\ai-work

claude

On first run it asks you the following, in order.

1. Pick a theme (colours) — choose with the arrow keys and press Enter. Any of them is fine.
2. Pick a login method — two options. (The wording shifts slightly between versions.)
   - Sign in with a subscription account -> for Claude Pro/Max subscribers. Most people pick this.
   - Sign in with a Console (API) account -> pay-as-you-go, for company or institutional accounts.
3. A browser opens automatically. Sign in to your Claude account and click "Authorize".
4. A long string (a code) appears on screen. Copy it, paste it into the terminal and press Enter.
5. It asks whether you trust the folder. Choose Yes, proceed.

The chat window appears. You only log in once; after this, typing claude starts you straight away.

Commands worth knowing

Inside the chat you can use special commands beginning with /.

4.3 Install OpenAI Codex CLI

Method A — dedicated installer (recommended)

irm https://chatgpt.com/codex/install.ps1 | iex

Method B — via npm

npm install -g @openai/codex

Method C — via winget

winget install --id OpenAI.Codex -e

Use only one of the three. Installing more than one way leaves you with tangled versions that are hard to diagnose later.

Check that it worked

Open a new terminal and type:

codex --version

Log in

codex login

A browser opens. Sign in with your ChatGPT account. ChatGPT Plus/Pro subscribers use the allowance included in their subscription, with nothing extra to pay.

To confirm you are logged in:

codex login status

Run it

Move into your working folder and type:

codex

4.4 Install Antigravity CLI (the successor to Gemini CLI)

First, some background — Gemini CLI is finished

On 19 May 2026 Google announced the end of individual-user support for Gemini CLI and the Gemini Code Assist IDE extensions, and service actually stopped on 18 June 2026. The successor is Antigravity CLI, whose terminal command is agy.

Who can still use Gemini CLI: holders of a Gemini Code Assist Standard/Enterprise licence, users of Code Assist for GitHub through Google Cloud, and holders of a paid Gemini API key. Individual users outside those categories cannot even log in.

If you already installed Gemini CLI, remove it first.

npm uninstall -g @google/gemini-cli

Install

Type this into an ordinary (non-administrator) PowerShell.

irm https://antigravity.google/cli/install.ps1 | iex

It installs to C:\Users\yourname\AppData\Local\agy\bin, and the install script registers it on your PATH for you.

If you use Command Prompt (cmd), do this instead.
curl -fsSL https://antigravity.google/cli/install.cmd -o install.cmd && install.cmd && del install.cmd

Check that it worked

Close the terminal completely, open a new one, then type:

agy --version

A version number means success. If it says agy cannot be found, nine times out of ten you did not open a new terminal. If that is not it, see section 9.1.

Log in

Move into your working folder and run:

agy

A browser opens automatically. Sign in with your Google account and grant permission; you return to the terminal with a chat window ready. Your credentials are stored in Windows Credential Manager, so from now on typing agy starts you straight away.

On first run it asks whether you trust the working folder (Workspace Trust). If you created the folder yourself, choose to trust it.

Commands worth knowing

Updating

Running the install script again upgrades you to the latest version.

irm https://antigravity.google/cli/install.ps1 | iex

Part 5. Your first session — actually giving it work

5.0 Learn these first — the five essential commands

Inside the chat window these five are effectively all you use. Know them and you will never be stuck.

Commands differ slightly between tools.
The five above are Claude Code's. Codex CLI and Antigravity CLI name some of them differently, and a few do not exist at all. Type /help in the chat window and you get the full list of commands and shortcuts that tool actually has. When in doubt, start with /help.

Your allowance resets on a fixed cycle

Paying a subscription does not buy unlimited use. The AI companies allocate an amount of usage (tokens) per fixed window, and refill it when that window rolls over. So hitting a limit is a wait, not a wall.

How to read this — "per 5 hours" means the window starts when you send your first instruction and covers the next five hours. Use up the allocation inside it and you wait for the next window to open. A weekly limit sits on top of that, so several heavy days in a row can hit the weekly cap even while your 5-hour allowance still has room.

No exact numbers here.
Allocations change often, and they differ by plan (Pro, Max, Plus and so on). Rather than memorising figures, get into the habit of checking with /usage.

Three ways to spend less of it

• /clear when the topic changes — a long conversation is re-read on every turn, which burns allowance fast. Clear it before starting a new subject.
• Use a lighter model for easy work — switch with /model. There is no reason to spend a heavy model on file tidying or simple conversions.
• Keep the working folder small — the fewer files the AI has to scan, the less it consumes. The "dedicated working folder" rule in Part 8 pays off here too.

Pick one of two paths

The road forks here. Either one is a fine place to start. There is no required order, and doing one does not stop you doing the other.

1. Practise in the terminal — just carry on reading from section 5.1. This is the better route for seeing what the tool actually does and what an approval screen looks like. It is also where you develop a sense of how to find the cause when something breaks.

2. Go straight to Orca — if the black window puts you off, or you want several jobs running from the start, take this route. Go to the Orca beginner's guide. Orca reuses the AI CLI you installed in Part 4, so if you finished Part 4 you are already set up.

The order I would suggest: spend thirty minutes on option 1, then move to option 2. Orca is ultimately a shell that launches several of the Part 4 tools, so it makes much more sense once you know what those tools are. That is a suggestion only — starting with option 2 causes no problems.

5.1 Prepare a working folder

An AI agent only handles files inside the folder you launched it from. That is a safety feature as much as anything.

Do not launch it on your Desktop or in C:\Users\yourname. That would let the AI rummage through every file on the machine.

Make a practice folder.

mkdir C:\ai-practice

cd C:\ai-practice

5.2 Make a practice file

There is nothing to do in an empty folder, so create one file to work with.

notepad notes.txt

Notepad asks whether to create the file. Choose Yes, write a few lines, save with Ctrl + S, and close it.

Example:

January 2026 revenue: 12,000
February 2026 revenue: 14,500
March 2026 revenue: 9,800

5.3 Run the AI and talk to it

claude

(If you installed Codex, type codex; for Antigravity CLI, agy.)

When the chat window appears, just talk to it in plain English.

Try this as your first message.

Check what files are in this folder and summarise their contents

The AI will look through the folder, read notes.txt, and summarise it.

Next, try this.

Turn the revenue data in notes.txt into a table, calculate the average and
the month-on-month change, and save it as revenue-summary.md

5.4 Understanding approval prompts — the most important part

When the AI is about to modify or create a file, you see a screen like this.

About to create file: revenue-summary.md

  1. Yes                         (allow once)
  2. Yes, and don't ask again    (allow from now on)
  3. No, tell Claude what to do  (refuse and redirect)

Choose with the arrow keys and press Enter.

Rules for beginners

• Read-only actions -> allow freely with 1. Yes
• Creating a new file -> 1. Yes
• Editing or deleting an existing file -> read it carefully before deciding
• Do not pick option 2 (allow from now on) at first. Save it for when you are comfortable.
• If something looks off, pick 3. No and say "don't do that, do X instead".

To stop a job midway, press Ctrl + C.

5.5 Quitting

Type this in the chat and you are done.

exit

Or press Ctrl + C twice.

5.6 Starting again next time

1. Open your working folder in File Explorer.
2. Right-click on empty space -> "Open in Terminal".
3. Type claude. (Or codex / agy.)

You are already logged in, so it starts immediately.

Part 6. Things worth trying

Instructions work better the more specific they are. Copy the examples below and adapt them.

6.1 File organisation

Rename every file in this folder to "date_title" format.
Use the file's creation date, formatted like 20260901.
Show me the planned list first, before renaming anything.

Look through the PDFs in this folder and build a table of
file name, title, page count and a one-line summary. Save it as index.md

6.2 Data work

Open data.csv, tell me how many rows have empty cells,
and suggest ways to handle them.
Do not modify the file yet.

Draw a bar chart of the monthly revenue in revenue.xlsx
and save it as a PNG. Put a title and axis labels on it.

6.3 Writing

Read draft.md and rewrite it in a dry, objective academic register.
Leave the original untouched and save the revision as draft-revised.md
Also summarise what you changed and why.

Write an email in three tones (formal / concise / friendly) covering the following.
Content: (describe the situation here)

6.4 Learning as you go

Explain what you just did in terms a beginner would understand.

Suggest five useful things I could do in this folder.

6.5 How to write good instructions

Four habits

1. Say what to do + how to do it + where to save the result, together.
2. For risky work, add "show me the plan first and run it only after I approve".
3. To protect the original, say explicitly "do not touch the original, work on a copy".
4. If you dislike the result, do not say "do it again" — say "change X to Y".

Part 7. Extending it — MCP servers and plugins

There are things an AI agent cannot do out of the box: open a Korean HWP document, check the text of a statute, pull national statistics. You add capabilities like these by installing an extension.

7.1 How MCP servers and plugins differ

Both widen the range of what the AI can do. Once installed, there are no commands to memorise — instruct it as you normally would and it picks up the extension when it needs to.

MCP is short for Model Context Protocol. It is the name of the standard an AI and an outside program use to talk, and a program built to that standard is called an MCP server.

7.2 The easiest way to install — have the AI do it

You do not have to hunt down the install command. Give it the repository address and ask.

Type this into the chat window.

Install https://github.com/chrisryugj/kordoc globally so I can use it.

The AI reads that repository's documentation, picks the install command that suits your machine, and runs it. Approval prompts appear along the way.

When it finishes, close the terminal completely and open a new one. Most newly installed extensions are only picked up after a restart.

Read the approval screen before you click.
Installing an extension puts a new program on your machine. The rules in Part 8 apply here too. Read which command it is about to run, and do not install from a repository you do not know. Once installed, an MCP server can reach the files and data the AI works with.

Sections 7.3 onward are for when you want to type the install commands yourself. If asking the AI worked, you can skip them.

7.3 kordoc — working with Korean HWP/HWPX documents

The most common sticking point in Korea. It reads, edits and generates HWP and HWPX files.

Method A — the setup wizard (recommended)

Type this into the terminal. It finds the AI clients on your machine and patches their configuration for you.

npx -y kordoc setup

Method B — as a Claude Code plugin

Inside the Claude Code chat window:

/plugin marketplace add chrisryugj/kordoc

/plugin install kordoc@kordoc

Try it — open a new terminal in a folder containing an HWPX file and ask:

Summarise report.hwpx and keep the tables intact

7.4 korean-law-mcp — Korean statutes and case law

It stops the AI inventing legal provisions. It queries the National Law Information Center (the Ministry of Government Legislation) directly and checks whether a citation actually exists.

Method A — as a Claude Code plugin (simplest)

/plugin marketplace add chrisryugj/korean-law-mcp

/plugin install korean-law@korean-law-marketplace

Method B — global install

npm install --ignore-scripts --omit=optional -g korean-law-mcp

The key you registered goes in the LAW_OC environment variable. Asking the AI is quicker than doing this by hand.

Set the LAW_OC environment variable so korean-law-mcp uses my API key. The key is: (your key)

Try it

Show me the text of Article 750 of the Civil Act and any recent amendments

7.5 korean-stats-mcp — Korean national statistics (KOSIS)

It pulls live data from Statistics Korea's KOSIS service. Every answer carries its source (the KOSIS table id), so it can go straight into a report.

Method — the Windows install script

Type this into the terminal.

irm https://raw.githubusercontent.com/chrisryugj/korean-stats-mcp/main/install.ps1 | iex

It asks for your KOSIS API key during installation. Paste in the key you registered.

Try it

Tabulate the population of Gwangju over the last ten years and cite the source

7.6 gptaku_plugins — a Claude Code plugin collection

Eighteen plugins that add capabilities to Claude Code, gathered in one marketplace.

Install — in the Claude Code chat window, one after another:

/plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git

/plugin install insane-search@gptaku-plugins

/reload-plugins

Once the marketplace is registered, pick up others with /plugin install <name>@gptaku-plugins. Start with one and get a feel for it.

7.7 Checking what you installed, and removing it

If something starts misbehaving, switch extensions off one at a time. With several installed it gets hard to tell which one is at fault. Install one at a time, and remove what you do not use.

Everything here is third-party open-source software. None of it is an official Anthropic, OpenAI or Google product. The repositories and install commands were checked in September 2026. These tools change, so if an install fails, check the repository's own documentation first.

Part 8. Using it safely — five rules

1. Only run it in a dedicated working folder

An AI agent treats the folder you launched it from, and everything below it, as its working scope.

• Do not: launch it on the Desktop, in C:\Users\yourname, or at the root of C:
• Do: create purpose-specific folders like C:\ai-work\project1 and launch inside them

2. Copy the originals first

The AI can edit a file wrongly, and you can approve the wrong thing.

Duplicate the whole folder before you start. It is the simplest and most reliable safeguard there is.

3. Keep sensitive material out of the working folder

File contents the AI reads are sent to a server to produce an answer.

Do not put these in your working folder:
- Files containing national ID numbers, bank account numbers or passwords
- Unpublished research data, manuscripts under review
- Company confidential material, contracts
- Lists containing other people's personal data

If your institution's rules forbid sending something outside, do not bring it into the folder in the first place.

4. Read the approval screen before you click

Hammering Yes to get through faster is a dangerous habit. Read what it is about to do, at least once.

Stop and check whenever you see these words in particular: delete, rm, overwrite, all, force.

5. Do not use auto-approve

Claude Code's --dangerously-skip-permissions, Codex CLI's --full-auto/--yolo, and Antigravity CLI's auto-approve setting all mean "stop asking and just do everything." They are as dangerous as they sound. Beginners should never use them.

Tip articles online sometimes recommend turning this on for convenience. That advice comes from people working in code repositories where a bad deletion is recoverable. Do not use it in a folder holding your original data.

Part 9. When something goes wrong

9.1 "The term ... is not recognized"

'claude' is not recognized as the name of a cmdlet, function, script file...

Try in this order

1. Close the terminal completely and open a new one. (Nine times out of ten this fixes it.)
2. If that fails, reboot.
3. Still failing? Check where it installed.

npm config get prefix

You get a path like C:\Users\yourname\AppData\Roaming\npm. Register it with Windows.

• Windows key -> search environment variables -> run "Edit environment variables for your account"
• Select Path in the upper list -> "Edit"
• "New" -> paste the path -> press "OK" on every window to close them
• Open a new terminal and try again

9.2 "...cannot be loaded because running scripts is disabled"

Windows is blocking script execution. The script-permission setting in Appendix E fixes it. One line into the terminal:

Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

Type Y and press Enter.

9.3 Non-English text shows as ??? or garbled characters

Temporary fix:

chcp 65001

Permanent fix: do the profile setup in Appendix E.

If that still fails, change a Windows setting.
- Windows key -> search region -> "Date, time & regional formatting" -> "Additional date, time & regional settings" -> "Region" -> "Administrative" tab -> "Change system locale"
- Tick "Beta: Use Unicode UTF-8 for worldwide language support" -> reboot

This can cause side effects in some older non-Unicode programs. Use it only when nothing else works.

9.4 Red ERR! errors during installation

If it is a permissions problem: reopen the terminal as administrator.
Windows key -> search Terminal -> right-click the entry -> "Run as administrator"

If it is a network problem: corporate and university firewalls block these installs often. Try again on a personal connection (phone tethering, for instance).

9.5 Trouble caused by spaces or non-English characters in a path

Paths like D:\My Research\2026 Data occasionally cause problems.

Fix: build working folders using letters, numbers and underscores only where you can.

• Good: C:\ai-work\project2026
• Avoid: C:\My Documents\2026 Research Data (final)

When you must use an existing path, wrap it in double quotes.

cd "D:\My Research\2026 Data"

9.6 The AI answers oddly, or the conversation gets tangled

In a long conversation the AI can get dragged along by earlier context.

• Type /clear -> resets the conversation (the same command in all three tools)
• Or exit and start again

Getting into the habit of pressing /clear whenever the topic changes noticeably improves answer quality.

9.7 Updating to the latest version

Update the same way you installed.

Do not use npm update -g. It only moves within the version range recorded at first install, so it often fails to reach the latest release. Use npm install with @latest.

If you installed Codex through winget, use winget upgrade --id OpenAI.Codex -e.

9.8 Removing them

npm uninstall -g @anthropic-ai/claude-code

npm uninstall -g @openai/codex

winget uninstall --id OpenAI.Codex -e

For dedicated installers, follow the uninstall instructions in each tool's official documentation.

If a retired Gemini CLI is still lying around, remove it like this.

npm uninstall -g @google/gemini-cli

Appendix A. Glossary

Appendix B. One-page command summary

Terminal basics

PowerShell is assumed. In the old Command Prompt (cmd), ls and pwd do not work (section 2.4).

Installation (once)

winget install --id OpenJS.NodeJS.LTS -e

winget install --id Git.Git -e

For AI agents, install only the one you need of the three.

irm https://claude.ai/install.ps1 | iex

irm https://chatgpt.com/codex/install.ps1 | iex

irm https://antigravity.google/cli/install.ps1 | iex

Verification

node --version

npm --version

git --version

claude --version

The everyday flow

1. Open your working folder in Explorer
2. Right-click empty space -> "Open in Terminal"
3. Type claude (or codex / agy)
4. Give instructions in plain English
5. Read the approval screen and choose
6. exit to quit

The five essential in-chat commands

Section 5.0 explains these, and the usage limits, in full.

Keyboard shortcuts

Appendix C. Summary for Mac users

This document is written for Windows. On a Mac, only these things differ; everything else is the same.

If you do not have Homebrew, install it first following the instructions at https://brew.sh.

For AI agents, use curl instead of irm ... | iex. Everything else is identical.

curl -fsSL https://claude.ai/install.sh | bash

curl -fsSL https://antigravity.google/cli/install.sh | bash

Install Codex CLI through Homebrew.

brew install --cask codex

The npm install -g ... route is exactly the same as on Windows.

Appendix D. Do you need WSL (Linux)?

Short answer: if you are the audience for this guide, just use Windows.

You will find articles online saying "Claude Code requires WSL". Most of them were written in 2024–2025, when it genuinely did not support Windows and WSL (Windows Subsystem for Linux, a feature that runs Linux inside Windows) was the only route. All three tools now ship Windows-native installers, and WSL is not needed.

Why plain Windows is the better choice

WSL is a net loss when your work is documents and data sitting in Windows folders. Reading and writing Windows drives (/mnt/c, /mnt/d) from WSL is measurably slow, and it becomes obvious once you are working across hundreds of files.

Do this instead

The way to reduce friction on Windows is not WSL — it is installing Git for Windows (section 3.3). When Claude Code detects Git Bash it uses that as its shell instead of PowerShell, and since most material online (and most of what the AI knows) is written in Linux-style commands, errors drop.

When WSL is the better option

Consider WSL if one of these applies to you. Most readers of this guide will not be in that position.

• You need Linux-only development tools (Docker, particular Python packages, and so on)
• You are building something that will run on a server and want your environment to match it
• You want to give the AI risky work while keeping it isolated from Windows itself

If you do go with WSL, keep your working files inside WSL (~/projects/...). Crossing back and forth to Windows drives is the slowest and most trouble-prone arrangement.

Appendix E. Optional settings — script permission and text encoding

You do not need either of these up front. They are optional, for when something actually breaks. Today's installers usually work without them.

E.1 Script permission

For security, Windows blocks external scripts by default. Lift that when an install or a run is blocked by it.

Type this into the terminal.

Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

When it asks whether you want to change the execution policy, type Y and press Enter.

Is that dangerous?
RemoteSigned is the middle setting: "run scripts I wrote myself, and run downloaded ones only if they are signed." It is the standard developers use, and it does not disable security. Because of -Scope CurrentUser it applies to your account only, not the whole machine.

To check the current setting without changing anything:

Get-ExecutionPolicy -Scope CurrentUser

E.2 Text encoding

Only needed if file names or output in a language other than English show as ??? or as garbled characters.

For a one-off fix, type one line into the terminal. It reverts when you close the window.

chcp 65001

To apply it automatically every time, put it in your profile file. Type this into the terminal.

notepad $PROFILE

Notepad opens. If it asks "Cannot find the file. Do you want to create a new file?", choose Yes.

Paste these three lines into Notepad.

[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
[Console]::InputEncoding  = [System.Text.Encoding]::UTF8
$PSDefaultParameterValues['*:Encoding'] = 'utf8'

Save with Ctrl + S and close Notepad. Close the terminal and open a new one.

This file ($PROFILE) runs automatically every time you start the terminal. Put it here once and you never touch it again.

If that still fails, consider the Windows system-locale change in section 9.3.

Wrapping up — what comes next

If you got this far, you have the fundamentals. Once you are comfortable, try these.

1. Put a rules file in your working folder — create a rules file saying "in this folder, work according to these rules" and the AI follows it without being told each time. Claude Code reads CLAUDE.md; Codex CLI and Antigravity CLI read AGENTS.md. In Claude Code, /init drafts one for you.

2. Save the instructions you reuse — keep the prompts for recurring work in a text file and paste them in.

3. Add extensions (MCP servers and plugins) — Korean documents, statutes, national statistics, Google Drive and more can be wired directly into the AI. Part 7 covers how to install them.

4. Compare the tools — giving the same job to Claude, Codex and Antigravity teaches you their different characters.

Official documentation

• Claude Code — https://code.claude.com/docs
• Codex CLI — https://developers.openai.com/codex/cli
• Antigravity CLI — https://antigravity.google/docs/cli/install

This document is written for Windows 11. Screens shift a little as versions move on, but the overall flow stays the same.