Command-line AI coding agents worth knowing

Claude Code, Codex, Copilot CLI, OpenCode, Aider… These AI agents write code for you from the terminal, the text window where you type commands, and the list gets longer every month. You’ll also hear them called “CLIs”, short for command-line interface.

Here are the nine worth knowing in September 2026, with what each one does best and what it costs. For this article, two of them, Codex and OpenCode, took on the same small exercise: writing a function in PHP (the language WordPress runs on) that turns a title into a clean web address. The French title “Cœur & âme” has to come out as coeur-ame, and “Ça coûte 100 €” as ca-coute-100. I checked every result myself, independently of the agent.

What is a coding agent?

A chatbot like ChatGPT replies with code, which you then copy over yourself. A coding agent does the job from start to finish. It opens your files, writes the code, runs the commands, looks at the result and starts again until the task is done. Because it lives in the terminal, it can also work with nobody in front of the screen, launched automatically by a script.

Three terms come up in the profiles below.

• MCP: a standard socket that plugs the agent into other tools, such as a database, GitHub or a web service, a bit like a USB port that accepts any device.
• Skills: instruction sheets the agent keeps in a drawer and only takes out when it needs them. Our guide to skills and the SKILL.md file explains how to write them.
• Sandbox: the enclosure that limits what the agent can read, change or send over the internet, covered in detail in the article on sandboxing in Claude Code and Codex.

This selection compares the tools, not the AI models running inside them. For those, see the head-to-head between GPT-6 Astra and Claude Fable 5.1.

1. Claude Code, Anthropic’s agent

Claude Code is the agent from Anthropic, the company behind the Claude models. It comes with the Claude Pro subscription and, since 1 September 2026, runs on Claude Fable 5.1 by default. Anthropic ships a new version almost every day, five in the first week of September alone, all listed on GitHub and in the release notes.

For a beginner, its best selling point is auto mode, switched on by default since 14 August on the Pro, Max and Team plans. Instead of asking for your permission every time it changes a file, Claude Code has its actions vetted by security checks that run in the background. It only reads a skill in full when it’s about to use it, and a built-in command spots the ones that never get used, as explained in our article on /skill-doctor. Its sandbox does exist, but it isn’t switched on out of the box.

2. Codex, OpenAI’s CLI

At OpenAI, the agent is called Codex, and it comes with the ChatGPT Plus subscription. Its source code is public, under the Apache 2.0 open source licence, which means anyone can read it, modify it and share it. It also handles skills and MCP connectors.

For the exercise, it worked with GPT-6 Astra, the model that arrived in Codex on 4 September. It got it right first time, in 56 seconds, and the tests I ran afterwards all passed. You can set Codex to three levels of freedom: read-only, writing limited to the project folder, or full access to the computer. The article on coding agent sandboxes tells the story of a very real loophole found in them.

On jobs lasting several hours, Codex also built a Dune 2000 prototype and then a small EarthBound-style role-playing game, two prototypes you can play in the browser.

3. GitHub Copilot CLI

Copilot CLI is the terminal version of GitHub Copilot, the coding assistant from GitHub, which is owned by Microsoft. It has been available to everyone since 25 February 2026, on Mac, Windows and Linux, and it comes already connected to GitHub through a built-in MCP connector.

Its big draw is the choice of model. You can switch from an Anthropic model to an OpenAI or Google one without changing tools, and Claude Fable 5.1 has been on the list since early September. You do need a paid Copilot subscription, though, Pro or higher, because the command-line version has no free plan.

4. Antigravity CLI, the replacement for Gemini CLI

If you’re looking for Gemini CLI, it has been closed to consumer accounts since 18 June 2026. Google is asking people to move to Antigravity CLI, launched on 19 May at its Google I/O conference, alongside the Google Antigravity desktop app.

Antigravity CLI works with an ordinary personal Google account, a Gemini Enterprise account or an API key, a kind of password reserved for programs. The most generous plan, AI Ultra, costs $100 a month and gives you five times as much usage as the Pro plan.

5. OpenCode, the one that coded without an account

OpenCode is an open source agent, released under the MIT licence by a company called Anomaly. It installs with a single command. With no account, no key and no bank card, it read the instructions, wrote the function, ran the tests itself and reported that everything passed, all in twenty seconds. I ran the tests again myself, with the same result.

The cost shown was zero. OpenCode used a model called Big Pickle, which the OpenCode Zen documentation describes as free for a limited time, to gather feedback. The other models in its catalogue are pay-as-you-go.

The flip side of this convenience is that OpenCode doesn’t ask for permission by default. During the exercise, it wrote a file and ran a command without asking anything. Our article dedicated to OpenCode walks through how to install and configure it.

6. Cursor, the editor that also has an agent

Cursor, from a company called Anysphere, made its name with its code editor and also offers a command-line agent. Its main focus is its cloud agents, which work on remote servers rather than on your computer. Since 2 September 2026, these agents can run on the servers of your choice, with AWS, Cloudflare, Vercel and others.

The Pro plan costs $20 a month, the Ultra plan $200, and team plans range from $40 to $120 per person. Every plan, starting with the first, includes MCP connectors, skills and hooks, which are triggers that launch an action at a precise moment.

7. Kiro, AWS’s agent

Amazon Web Services (AWS), the hosting giant, also has a command-line agent, called Kiro. It comes with its own vocabulary. “Powers” play the role of skills, and “Steering” brings together rule files that guide its work. Since August, you can dictate your instructions to it by voice, or send it off to work in an isolated space hosted by AWS.

The free plan gives you 50 credits a month, with open models and Claude Sonnet 4.5. The Pro plan, at $20 a month, gives you 1,000, and each extra credit costs 4 cents.

8. Aider, the terminal veteran

Aider is the oldest agent on the list and the one most devoted to the terminal, with no app and no cloud. This open source project, which has passed 44,000 stars on GitHub (the equivalent of “likes”), works with almost any model, including ones that run on your own computer.

Aider records every change in the project’s history, so you can go back at any time. It’s free, and you only pay for the model you pick. Unlike the other agents on this list, it still doesn’t support MCP natively.

9. Cline and Roo Code

Cline mainly lives inside code editors (VS Code and JetBrains), but there’s also a command-line version, which can run unattended inside a script. Open source and backed by 67,600 stars on GitHub, it accepts more than 200 models, from those made by Anthropic, OpenAI or Google to local models through Ollama, a piece of software that runs an AI on your computer.

Roo Code is a spin-off of Cline, known for a while as Roo Cline, and it has been installed more than 50,000 times. Both are free, and you only pay for the model. You pick one or the other based on the extensions on offer and how quickly new features arrive.

OpenClaw and Hermes Agent, two assistants that also code

These two names often come up in the same conversations, even though they aren’t coding agents in the strict sense.

OpenClaw

OpenClaw connects your messaging apps (Discord, Signal, Telegram or WhatsApp) to an AI agent that can, among other things, write code. Created by the Austrian developer Peter Steinberger, the project changed its name twice in early 2026. It became Moltbot on 27 January, at the request of Anthropic, which felt the original name was too close to Claude, then OpenClaw three days later. Steinberger joined OpenAI on 14 February and handed the project over to a non-profit foundation. Our article on installing it with Docker explains how to set it up properly and keep it secure.

Hermes Agent

Hermes Agent, from Nous Research, is an open source personal assistant that answers you in the terminal, on around twenty messaging apps and in a desktop app, with more than 300 models to choose from. When there’s code to write, it hands the job to OpenCode, which it runs inside a sandbox. A separate article tries out installing it with Docker.

The comparison at a glance

Is there a best AI for coding?

No, not in September 2026. The best AI for coding depends on what you already pay for, your budget and the code you plan to hand over to it.

• You’re just starting out. Go for Claude Code or Codex. Claude Code’s auto mode and the very detailed documentation of both mean fewer decisions for you to make.
• You already have a subscription. Start with the agent it includes: Codex with ChatGPT Plus, Claude Code with Claude Pro, Copilot CLI with Copilot, Antigravity CLI with a Google account.
• You’re looking for a free AI for coding. OpenCode finished the exercise without a card or a key, but its free model is temporary, so don’t build anything that depends on it. Aider and Cline are free too, and you only pay for the model. A model installed on your own computer costs nothing to run.
• You work in a team. Claude Code, Cursor and Copilot CLI let you enforce the same MCP connectors across the whole team, which matters more than a model ranking that changes every month.
• Your code is sensitive. None of these agents keeps its work contained by default, apart from the cloud modes in Kiro and Cursor. Read the article on sandboxing coding agents before you connect one to a project that matters.

What to remember

• There’s no single best AI for coding. The right choice depends on your subscription, your budget and how sensitive your code is.
• Codex passed the exercise first time, in under a minute.
• OpenCode passed it with no account and no bank card, thanks to a model that’s free for a limited time.
• Gemini CLI has been closed to the general public since 18 June 2026, and Antigravity CLI is taking its place.
• Aider is the only one on the list that doesn’t support MCP natively.
• None of these agents runs in a sandbox by default, so set one up before you trust an agent with an important project.