A beginner's guide · 2026

What is an AI coding agent, and which one should you try first?

There are dozens of AI coding agents now — the ecosystem is moving fast. Here are four popular ones to start with. Pick one and we'll get you running in fifteen minutes.

Start with the 4-step guide↓Or skip to compare these four picks

§ 01The basics

An agent is a chatbot that does the work for you.

You give it a goal — fix this bug, build me a script, set up a project — and it writes, edits, and runs code in your files and your terminal until the goal is met.

The difference from ChatGPT is that an agent doesn't stop at suggesting code. It opens the file, makes the change, runs the test, reads the error, and tries again. You stay in the loop — but you're directing, not typing.

People use them to build apps, fix bugs, write one-off scripts, automate boring chores, and learn programming faster than going it alone.

How a run works

You

Type a plain-English goal. “Fix the broken login button.” Hit enter.

Agent reads

It opens the relevant files, traces the code, and forms a plan.

Agent writes & runs

Edits the files, runs the tests, reads the output, retries until it works.

Working code

You read what changed, approve it, and you're done. The whole loop takes minutes, not hours.

The trick nobody tells you

You don't need to know what to ask. You just need to ask.

Once an agent is installed, the very first thing you can ask is “this is my first time — what can you do, and how do I use you?” Agents are built to explain themselves. They'll walk you through. The hard part is the install, not knowing the magic words.

→ your first prompt is the easy part → the agent does the rest

§ 02Quick router

Not sure which of the four to start with? Pick the closest match.

Tap a row → jumps to comparison

01If you want a hosted, fully-managed experience with sensible defaults out of the box,→Claude Code
02If you already use ChatGPT (any plan, including Free) and want to use what you've got,→Codex
03If you want an agent that learns skills and gets better the more you use it,→Hermes
04If you want to run agents on your own machine with your own model,→OpenClaw

§ 03The four at a glance

Side-by-side, in plain language.

Four things that actually matter when you’re deciding which to start with. Dozens of other agents exist; these four are popular starting points.

Agent

Made by

Model

Cost to start

Setup

Claude Code

Anthropic

Claude only

Free trial, then API or Pro plan

Easy

Codex

OpenAI

OpenAI only

Free with any ChatGPT plan

Easy

Hermes

Nous Research

Bring-your-own

Free, plus your model's API

Medium

OpenClaw

Bring-your-own

Free, plus your model's API

Hard

Claude Code

Made by: Anthropic
Model: Claude only
Cost: Free trial, then API or Pro plan
Setup: Easy

Codex

Made by: OpenAI
Model: OpenAI only
Cost: Free with any ChatGPT plan
Setup: Easy

Hermes

Made by: Nous Research
Model: Bring-your-own
Cost: Free, plus your model's API
Setup: Medium

OpenClaw

Made by: OpenClaw
Model: Bring-your-own
Cost: Free, plus your model's API
Setup: Hard

Claude Code

docs.claude.com ↗

Anthropic's official agent. Sensible defaults, useful built-ins, and Claude under the hood. The trade-off: it's Claude only — if you want to swap in a different model later, you can't.

Codex

openai.com/codex ↗

OpenAI's agent, included with any ChatGPT plan (Free, Go, Plus, Pro, Business, Enterprise). Strong cloud delegation — give it a task, walk away, come back when it's done. OpenAI only, and it lives in the browser more than your terminal.

Hermes

hermes-agent.nousresearch.com ↗

From Nous Research. Hermes learns skills as you use it — sessions stack, so it gets faster at the things you do often. Bring-your-own model. The most novel of the four; also the most experimental.

OpenClaw

docs.openclaw.ai ↗

Self-hosted, runs anywhere — terminal, WhatsApp, Telegram, Slack. Bring your own model and your own keys. The most flexible of the four, and the most setup work — onboarding has the most moving pieces. Best if you've outgrown the hosted options or want full control of your data.

§ 04Before you start

A short checklist so you don't hit a wall in step one.

If anything here is unfamiliar, that's okay — every agent's docs cover the install in detail. This list is what universally needs to be in place.

A computer with a terminal

Mac, Linux, or Windows with WSL. Anything from the last five years works.

Node.js or Python

Depending on the agent. Step 1 of the guide tells you which one and how to install it.

A free account with the provider

Anthropic, OpenAI, or — for the self-hosted ones — an API key from any model vendor.

A code editor

VS Code, Cursor, or whatever you already use. The agent runs in your terminal — the editor is just for opening files and seeing what the agent wrote.

Fifteen minutes

That's the realistic floor for install + auth + first prompt + first small win.

§ 05The 4-step guide

Get your first agent running in fifteen minutes.

Pick an agent →

Step 01

Install

~3 min

Open your terminal and paste the command below. New to terminals? Pick your OS:

Mac

⌘ Space → "terminal"

Windows

⊞ → "cmd" or "powershell"

Don’t have Node.js yet? Install it first.

nodejs.org↗

All platforms· recommended · uses npm

~/projects · zsh

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

Verify it workedtakes 5 seconds

When the install finishes, your terminal goes back to the prompt. Check that the new claude command is recognized:

~/projects · zsh

$ claude --version

You should see a version number printed back. If you see command not found, your terminal hasn’t picked up the new command yet — close the terminal and open a fresh one, then run the version check again. If it’s still missing, the install didn’t complete: re-run the install command above.

Other install optionspick the one for your setup

Permission error?· Mac / Linux only

~/projects · zsh

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

Mac / Linux· no Node.js — Anthropic's native installer

~/projects · zsh

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

Windows· no Node.js — cmd / Command Prompt

~/projects · zsh

$ curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

Windows· no Node.js — PowerShell

~/projects · zsh

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

Docs →docs.claude.com

Step 02

Authenticate

~2 min

Log in or paste your API key. Most of these use a browser-based login by default — type the command and a tab opens. If you don't have an account yet, signup is free.

claude · auth

$ claude login

Docs →docs.claude.com → setup

Step 03

Run your first prompt

~3 min

After step 2, you should see a > prompt — that’s the agent running and waiting for you. If you got dropped back to your shell instead, type claude and hit enter to start it up again.

First, ask the agent to set up a fresh empty folder for our work, so it only touches files in one safe place. Type these at the > prompt — one at a time, hitting enter after each. The agent will run them as shell commands for you (most agents ask permission first):

claude · session

> mkdir agent-test

mkdir is short for “make directory” (a directory is just a folder). This creates a new folder called agent-test wherever you currently are — by default that’s your home folder.

claude · session

> cd agent-test

cd is short for “change directory” — it moves into a folder. The agent is now working inside agent-test, so any files it creates land there.

Now your first real prompt — let the agent introduce itself. Paste this and hit enter:

claude · session

> Hey, this is my first time using a coding agent. Can you guide me step by step on what you can do?

The agent will explain its commands, ask what you want to build, and walk you through. Anything goes from here.

Step 04

Try a real project

~7 min

Pick something small that makes you smile when it works. The point is to feel the loop end-to-end — type a goal, watch it write code, open the result, fix things. Once you’ve built one fun thing, you’ll know what’s possible.

Beginner tipno install needed

Ask for things as a single HTML file. To run it, you double-click the file in Finder (Mac) or File Explorer (Windows) and it opens in your browser — no Python, no Node, no extra software. The HTML file the agent makes will be inside your agent-test folder.

claude · session

> Build me a tamagotchi-style virtual pet as a single HTML file I can open in my browser by double-clicking. No install or extra software needed.

Other ideas (still single HTML file): a sticky-note board · a habit tracker · a retro snake game · a daily-compliment generator

+ Bonus

Add a skill

~30 sec

Skills are reusable capability packs you install with a single prompt — they teach your agent a whole new domain in one go. Try this one to see how easy it is. Paste it at the > prompt:

claude · skills

> install the health skills from health.directivecreator.com

Your agent will fetch the skills, install them, and tell you what it can now help with. From here, ask it anything health-related — plan a week of meals around better sleep, build a workout routine, figure out why you crash at 3pm — and you’ll get answers with depth, not generic chatbot replies.

Skills compose. Install one, then another, then another. Each one expands what your agent can do for you.

§ 06Common questions

The questions everyone asks at minute zero.

Tap any row to expand. If your question isn't here, follow on X — most updates land there first.

No — but a little helps. You'll move faster if you can read what the agent writes and notice when something looks wrong. Many beginners use these tools as their first programming teacher: you ask for something, watch the agent build it, read what it wrote, and learn how the pieces fit together.

§ 07 Stay current

These tools change fast. Follow along.

Follow @socialure on X↗

No email. No newsletter. Just the feed.