12  12. The AI Coding Tool Landscape (Fall 2026)

There are dozens of AI coding tools, and their names, owners, and prices change every few months. This page reflects the landscape as of October 2026. To stay useful after names change, it is organized by category (the kind of tool) rather than by brand. The book focuses on Claude Code, but most of what you learn transfers to the other tools.

12.1 Categories of AI coding tools

As of Oct 6, 2026 · Official docs: Platforms and integrations

Category What it is Examples (October 2026)
Terminal agents (CLIs) An agent you run in a terminal, inside your project folder; it reads and edits files and runs commands Claude Code, OpenAI Codex CLI, Google Antigravity CLI, GitHub Copilot CLI, Cursor CLI, Grok Build
AI editors and agent apps A code editor with AI built in, or a desktop app for running and reviewing several agents at once Cursor, Google Antigravity, Devin Desktop, Kiro; the GitHub Copilot app, Codex in the ChatGPT desktop app, the Claude desktop app
Editor extensions A plug-in that adds an AI agent to an editor you already use, such as VS Code or a JetBrains IDE GitHub Copilot, Claude Code for VS Code and JetBrains, the Codex extension, Cline
Cloud agents (also called async agents) You hand off a task; the agent works on a cloud computer and usually returns a pull request (a proposed change for you to review) Codex cloud, Google Jules, Copilot cloud agent, Cursor Cloud Agents, Devin, Claude Code on the web
Prompt-to-app builders A website that turns a plain-English description into a working app, often hosted for you; you may never see the code Replit, Lovable, Bolt.new, v0, Google AI Studio
Open-source agents Agents whose source code is public; the agent is free, and you connect (and pay for) a model of your choice OpenCode, Cline, Goose, Aider, OpenHands; Codex CLI, Gemini CLI, and Grok Build are open source too
Open-weight models Not tools but models whose trained weights anyone can download and run; many tools let you plug them in DeepSeek, Qwen (Alibaba), Kimi (Moonshot AI), GLM (Z.ai), MiniMax, Devstral (Mistral); from the US, OpenAI’s gpt-oss and Google’s Gemma

Two patterns run through the table:

  • One vendor, many surfaces. Most big vendors ship the same agent in several of these categories, under one account and one usage allowance. Each of these forms is called a surface.
  • The tool is not the model. The model (Claude, GPT, Gemini, Grok, …) writes the code; the tool gives the model its tools and runs the loop (see AI-Assisted Development). Many tools (GitHub Copilot, Cursor, VS Code, Kiro, Zed, OpenCode) let you choose models from several companies, and model names change every few months.

12.2 The major vendors

12.2.1 Anthropic: Claude Code

As of Oct 6, 2026 · Official docs: Claude Code overview · Platforms and integrations

Claude Code is the agent this book uses. Its most complete form is the terminal tool; it also runs in VS Code (the extension also installs in Cursor), JetBrains IDEs, the Claude desktop app, and the cloud at claude.ai/code, all sharing one engine, so a project’s CLAUDE.md, settings, and MCP servers work in each. It comes with paid Claude plans (not the free plan) or a pay-per-use Claude Console account; see Claude Models, Plans, and Usage Limits.

12.2.2 OpenAI: Codex

Codex is included in ChatGPT plans, with all of its surfaces on paid plans: an open-source terminal tool (Codex CLI), an extension for VS Code-based editors, cloud tasks that open pull requests on GitHub, and Codex inside the ChatGPT desktop app (the separate Codex app was folded into it in July 2026). It also works with a pay-per-use API key. openai.com/codex · documentation

12.2.3 Google: Antigravity, Jules, and AI Studio

Antigravity is Google’s main coding platform: a desktop app for managing several agents in parallel, an AI editor, a terminal tool, and editor extensions, with a free tier for individuals. Jules is a cloud agent that turns a task on your GitHub repository into a pull request, and AI Studio includes a prompt-to-app builder. Older tutorials may mention Gemini CLI: since June 2026 it serves only organizations and paid API keys, and individuals were moved to Antigravity CLI (geminicli.com). antigravity.google · jules.google · aistudio.google.com

12.2.4 GitHub and Microsoft: Copilot and VS Code

GitHub Copilot (GitHub is owned by Microsoft) began in 2021 as autocomplete and is now a full agent: in editors, in the terminal (Copilot CLI), as a desktop app, and as a cloud agent on GitHub.com (formerly the “Copilot coding agent”) that opens pull requests. It is built into GitHub’s issue and pull-request workflow, offers models from several companies, and has a free Copilot Student plan. Microsoft’s VS Code also has agents built in and lets you choose which agent runs: Copilot, Claude, or Codex. github.com/features/copilot · VS Code agents

12.2.5 Cursor

Cursor is an AI-first code editor built on VS Code by the company Anysphere, known for fast “Tab” completions and an agent with a Plan Mode. It also offers Cursor CLI, Cloud Agents, Bugbot (an AI pull-request reviewer), and its own Composer models alongside other companies’ models. cursor.com

12.2.6 xAI: Grok Build

xAI (which now brands itself SpaceXAI) makes the Grok models and Grok Build, an open-source terminal agent (the grok command) that reads both AGENTS.md and CLAUDE.md. Grok models are also available inside Cursor and GitHub Copilot. Grok Build docs

12.2.7 Others worth knowing

  • Devin Desktop (Cognition; the editor formerly called Windsurf): an AI editor, a CLI, and the Devin cloud agent. devin.ai/desktop
  • Kiro (Amazon Web Services): an AI editor, CLI, and web version built around spec-driven development; AWS points users of its older Amazon Q Developer plug-ins here. kiro.dev
  • Junie (JetBrains): the agent for JetBrains IDEs, also a CLI. jetbrains.com/junie
  • Zed: an open-source editor that can host other companies’ agents. zed.dev
  • App builders: Replit, Lovable, Bolt.new, Vercel’s v0.
  • Open-source agents: OpenCode, Cline, Goose, Aider.

Chinese labs (DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, Z.ai’s GLM, and MiniMax) release strong open-weight models and sell low-cost access to them. Some US tool makers build on them: according to the companies’ own announcements, Cursor’s Composer models and Cognition’s SWE models start from Moonshot’s Kimi models. Tools such as Kiro and Cursor also offer some of these models directly. The result is downward pressure on prices and more “bring your own model” choice.

Two cautions: licenses vary (some are permissive, such as MIT or Apache 2.0; others have custom terms), and before you send code to any third-party service, especially one in another country, check your school’s or employer’s data rules. Open-weight versus closed models are explained in What AI Can Do Today.

12.3 How to choose

If you want to… Consider Trade-off
See every command and file change, and automate tasks with scripts A terminal agent You need basic command-line skills (this book teaches them)
Watch the code change in an editor as you work An AI editor or an editor extension An AI editor means switching editors; an extension keeps yours
Hand off a well-defined task and review the result later A cloud agent Less steering while it works; usually needs your code on GitHub
Get a clickable prototype quickly, without coding A prompt-to-app builder Less control over the code; check whether you can export it
Use any model, including cheap or local ones An open-source agent More setup; quality depends on the model you connect

A few more guidelines:

  • Pick the surface, then the brand. Most vendors offer several surfaces, so first decide how you want to work.
  • Budget realistically. As of October 2026, free tiers with tight limits and individual plans of about $20 a month are common, and heavy agent use is metered; check current pricing. Look for student offers, such as GitHub’s free Copilot Student plan or a free year of Google AI Pro for eligible US college students (redeem by December 31, 2026; it renews unless you cancel; details).
  • Check the data rules (your school’s or employer’s policy, the tool’s privacy settings) before you give a tool code from a client, an employer, or a class.
  • Expect names to change. Tools get renamed, merged, or retired. Trust the vendor’s official site and changelog over old tutorials or an AI’s memory.
  • Keep your know-how portable: project instructions and specs in plain-text files in Git work with any tool.

12.4 Why this book uses Claude Code

As of Oct 6, 2026 · Official docs: Claude Code overview · Platforms and integrations

A class works best when everyone uses the same tool. This book chose Claude Code for teaching reasons:

  • You see what the agent does. In the terminal, each file edit and command is shown as it happens, so learning Claude Code goes hand in hand with learning Bash and Git.
  • One engine, many surfaces. What you learn in the terminal also works in the editor extensions, the desktop app, and the cloud.
  • It has every building block we teach (instruction files, permission modes, MCP servers, skills, hooks, subagents, plugins), with detailed official documentation and a public changelog.

This is a teaching choice, not a claim that one tool is best for everyone. Many of the tools above are capable and widely used, and your skills carry over.

12.5 Concepts that transfer

As of Oct 6, 2026 · Official docs: CLAUDE.md and AGENTS.md · Commands

Coding agents have converged on the same building blocks. Learn them once in Claude Code and you will recognize them everywhere:

Concept In Claude Code In other tools
The agent loop: read, act, check, repeat The agentic loop Every coding agent
A project instruction file CLAUDE.md; Claude Code also reads AGENTS.md when a project has no CLAUDE.md AGENTS.md is read by Codex, Cursor, Copilot, Antigravity, Jules, Kiro, Junie, Devin, Zed, OpenCode, and Grok Build. Copilot and Grok Build also read CLAUDE.md.
Connecting outside services MCP servers Every major coding agent supports MCP
Deciding what may run without asking Permission modes and rules Look for the matching approval settings in any agent that runs commands
Planning before building Plan mode For example, Cursor’s Plan Mode and Kiro’s specs
Reusable helpers Skills, subagents, hooks, plugins Similar features in Codex, Cursor, Antigravity, VS Code, and Grok Build
Saving and reviewing work Git: diffs, commits, branches, pull requests The same Git everywhere

AGENTS.md is a plain Markdown file of project instructions, an open format stewarded by the Agentic AI Foundation under the Linux Foundation (agents.md). MCP (Model Context Protocol) is the open standard for connecting AI tools to outside tools and data (modelcontextprotocol.io). So one instruction file can serve several tools: keep it in AGENTS.md, and if you also need a CLAUDE.md, start that file with the line @AGENTS.md to import it (CLAUDE.md and Memory). If you have already set up Codex, Gemini CLI, or Cursor, Claude Code’s /import command brings their instruction files, MCP servers, commands, subagents, and skills over; typing /import cursor --dry-run in Claude Code previews an import from Cursor without changing anything.

12.6 Key takeaways

  • Think in categories: terminal agents, AI editors and agent apps, editor extensions, cloud agents, prompt-to-app builders, open-source agents, and open-weight models.
  • Most vendors offer several surfaces, and the tool is separate from the model.
  • Choose by the kind of work, your budget, and your data rules; names and prices change, so check official sites.
  • The agent loop, instruction files, MCP, permissions, and Git transfer from Claude Code to other tools.

Practice Questions

Easy

Question 1 · Matching

Match each category with its description.

Category Description
1. Terminal agent A. A website that turns a plain-English description into a working, often hosted, app
2. Editor extension B. Works on a computer in the cloud and usually delivers a pull request
3. Cloud agent C. Runs in a terminal inside your project folder, editing files and running commands
4. Prompt-to-app builder D. A model whose trained weights anyone can download and run
5. Open-weight model E. A plug-in that adds an AI agent to an editor you already use

1 → C, 2 → E, 3 → B, 4 → A, 5 → D.

Question 2 · Matching

Match each product with the company that makes it.

Product Company
1. Codex A. Anthropic
2. Antigravity and Jules B. GitHub (owned by Microsoft)
3. Copilot C. OpenAI
4. Grok Build D. Google
5. Claude Code E. xAI

1 → C, 2 → D, 3 → B, 4 → E, 5 → A.

Question 3 · True or false

Each big vendor offers just one kind of tool, so choosing a vendor also decides whether you work in a terminal, in an editor, or in the cloud.

False. Most big vendors (Anthropic, OpenAI, Google, GitHub, Cursor) ship the same agent on several surfaces: a terminal tool, an editor or extension, a desktop app, and a cloud agent, usually under one account. You choose the vendor and the surface separately.

Question 4 · Short answer

What is the difference between an AI coding tool and an AI model? Give one example of each.

The model is the AI that reads and writes the code, for example Claude, GPT, Gemini, or Grok. The tool is the program around the model that gives it tools (read files, run commands, …) and runs the loop, for example Claude Code, Cursor, or GitHub Copilot. Many tools let you choose among models from several companies.

Question 5 · Multiple choice

What is AGENTS.md?

  1. A list of AI companies and their agents, published by the Linux Foundation
  2. A plain Markdown file of project instructions that many coding agents read
  3. A settings file that only Claude Code understands
  4. A log file in which an agent records every command it runs

b. AGENTS.md is an open format for project instructions, read by Codex, Cursor, Copilot, and many other agents. Option a mixes up the file with its steward: the format is looked after by a foundation under the Linux Foundation, but the file is not a list of companies. Option c is wrong because many tools read it (Claude Code’s own file is CLAUDE.md, and Claude Code reads AGENTS.md when a project has no CLAUDE.md). Option d is wrong: it is instructions for the agent, not a log from it.

Question 6 · Short answer

What does MCP stand for, and what is it for?

Model Context Protocol. It is an open standard for connecting AI tools to outside tools, services, and data, such as GitHub, a database, or Google Drive. Because every major coding agent supports it, the same MCP server can work with different agents.

Medium

Question 7 · Scenario

Which category of tool fits each situation best?

  1. A marketing student who has never coded wants a clickable prototype of a club website for tomorrow’s meeting.
  2. A team wants an agent to work through several GitHub issues overnight and have pull requests waiting in the morning.
  3. A student wants to see every command the AI runs and learn the command line at the same time.
  4. A developer wants a free agent that uses a model running on her own laptop.
  1. A prompt-to-app builder: fastest way to something clickable with no coding. Fine for a prototype, even though she gets less control over the code.
  2. A cloud agent: it works in the cloud while the team is away and delivers pull requests to review.
  3. A terminal agent: every command and file edit is shown in the terminal.
  4. An open-source agent connected to an open-weight model that runs locally.

Question 8 · Multiple choice

Which statement is correct?

  1. An open-source agent and an open-weight model are the same thing.
  2. An open-source agent’s code is public, but you still connect it to a model, which may be an open-weight model or a commercial one.
  3. Open-weight models can only be used in tools made by the same company.
  4. Open-source agents work only with models from the company that wrote them.

b. An agent is a program (the tool); an open-weight model is a model whose weights are published. Option a confuses the two. Option c is wrong: mainstream tools such as Kiro and Cursor offer open-weight models from other companies, and anyone can download and run them. Option d is wrong: open-source agents such as OpenCode connect to many different model providers.

Question 9 · Scenario

Your team’s repository has an AGENTS.md file and no CLAUDE.md.

  1. What does Claude Code use as its project instructions?
  2. A teammate adds a CLAUDE.md with a few Claude-only notes. What changes, by default, and how can the team keep using the shared AGENTS.md?

  1. It reads the AGENTS.md file.

  2. Once a CLAUDE.md exists, Claude Code by default reads the CLAUDE.md files only, so it would stop reading AGENTS.md. The fix is to start CLAUDE.md with the line @AGENTS.md, which imports the shared file; the Claude-only notes go below it. (Claude Code also has a “Project instructions” setting in /config that can make it read both files.)

Question 10 · Write the command

Last semester you set up Cursor with project rules and MCP servers. Which command, typed inside a Claude Code session, previews what Claude Code would bring over from that setup without changing anything?

/import cursor --dry-run

Without --dry-run, /import cursor performs the import. The same command also works with codex and gemini. It is a Claude Code command, typed at the Claude Code prompt, not in Bash.

Question 11 · Short answer

Why does this page group tools into categories (terminal agent, cloud agent, and so on) instead of simply listing brand names?

Brand names, owners, and prices change often: tools get renamed, merged, or retired. The categories describe how you work with a tool, and they stay stable. Once you know the categories, you can place any new tool quickly, compare like with like, and choose the kind of tool before the brand.

Question 12 · Multiple choice

Which of these is the weakest reason to choose an AI coding tool?

  1. It fits the way you want to work (terminal, editor, or cloud).
  2. Your school’s or employer’s data policy allows it.
  3. It was the most-hyped tool on social media this week.
  4. It supports AGENTS.md and MCP, so your setup carries over.

c. Hype says nothing about whether a tool fits your work, your budget, or your data rules. Options a, b, and d are all solid reasons: the right surface, permission to use it, and a setup that stays portable.

Question 13 · True or false

If you learn Claude Code, very little of that knowledge will help you with GitHub Copilot or Codex.

False. The agent loop, project instruction files (AGENTS.md / CLAUDE.md), MCP, approval settings, planning before building, and all of your Git and command-line skills carry over. What changes are the names of commands, menus, and settings.

Hard

Question 14 · Compare

For a semester project that you will keep improving and must hand in with its code, compare a prompt-to-app builder with a terminal agent. Give at least two advantages of each.

Prompt-to-app builder: the fastest start, nothing to install, hosting often included, and great for an early demo, even for non-programmers.

Terminal agent: full control over the code and files in your own folder; works with Git (history, review, undo); deploy anywhere; you see every command and change, so you can understand and explain the code; easier to switch tools later.

Trade-offs: a builder gives you less control and may be hard to export or customize; a terminal agent needs command-line skills and some setup. A sensible plan for this project: prototype quickly if you like, then move the code into a Git repository and continue with an agent.

Question 15 · Scenario

Your summer internship uses GitHub Copilot, not Claude Code. Which skills from this course carry over directly, and what will you need to learn fresh?

Carries over: Bash and Git (including pull requests); how the web works; writing specs; the plan → build → verify → review → commit workflow; context engineering; instruction files (Copilot reads AGENTS.md, and even CLAUDE.md); MCP; the idea of approval settings; planning before building.

Learn fresh: Copilot’s own menus, commands, and settings; how its cloud agent is given work on GitHub; its plans and usage limits; and the company’s own rules for AI tools.

Question 16 · Scenario

A friend suggests a very cheap AI model service, run by a company in another country, for your internship’s code. What should you check before using it?

  • Your employer’s policy: is sending source code to this service allowed? Ask before you try it.
  • Data handling: does the service store your code, and for how long? Is it used for training? Which country’s laws apply?
  • Sensitive content: client data, secrets, or API keys in the code.
  • The model’s license terms.
  • Whether your approved tools can even be pointed at it, and who pays.

A low price is only one factor; data rules come first.

Question 17 · Long answer

Explain in your own words why this book uses Claude Code. Then give one legitimate reason a different team might choose a different tool.

Why this book uses it: a class needs one shared tool. In the terminal you see every command and edit, which fits with learning Bash and Git. The same engine runs on many surfaces (terminal, editors, desktop, cloud), and it has all the building blocks the course teaches (instruction files, permissions, MCP, skills, hooks, subagents, plugins), with detailed docs and a changelog.

Legitimate reasons for another choice (any one): the team lives in GitHub issues and pull requests and wants Copilot’s cloud agent; the company already pays for ChatGPT, which includes Codex; the developers prefer working in an AI editor such as Cursor or Antigravity; they use JetBrains IDEs (Junie); they need specific models, including open-weight or local ones (an open-source agent); or their data policy approves a different tool. A good answer stays neutral: the choice depends on fit, cost, and rules.

Question 18 · Scenario

You find a tutorial from 2025 that tells you to download a coding tool, but the tool’s website now redirects to a product with a different name. What should you do?

Don’t follow the old steps blindly. Go to the official site of the current product and look for its documentation and changelog or announcement about the change. Check whether the features and settings from the tutorial still exist and whether your account or plan carries over, and prefer recent official instructions. Be careful with AI answers about it too: a model’s knowledge may end before the change. (This happened in 2026, for example, when the Windsurf editor became Devin Desktop.)

Question 19 · Short answer

“Separate the tool from the model.” Describe one way a change you don’t control could affect your project, and two habits that protect you.

Example changes: the model you rely on is retired or replaced by a newer version that behaves differently; your tool stops offering a model; a business deal changes which models or plans a tool offers; prices or usage limits change.

Protective habits (any two): keep your instructions and specs in plain-text files in Git (AGENTS.md / CLAUDE.md, SPEC.md), so they work with any tool and any model; have tests, so you can check quickly whether a new model still does the job; avoid depending on one tool’s special features for critical work; read the changelogs of the tools you use.

Question 20 · Scenario

Your project team uses Claude Code and Codex on the same repository. Describe a setup so that both agents follow the same project rules and can reach the same database tool through MCP.

  • Keep the shared rules in AGENTS.md, which Codex reads, and which Claude Code reads when there is no CLAUDE.md.
  • If Claude-only notes are needed, add a CLAUDE.md whose first line is @AGENTS.md, with the extra notes below it, so the rules live in one place.
  • Add the same MCP server in each tool. Both support MCP, but each tool has its own configuration, so set it up once per tool.
  • Commit the instruction files to Git so the whole team shares them. Keep passwords and API keys out of these files and out of Git.
  • Both agents then follow the same workflow: small steps, verified, reviewed, and committed.

© 2025-2026 Y. Rosenthal. All rights reserved.