What is an AI coding agent?
An autocomplete tool suggests the next line. A coding agent takes a whole task, like "fix this failing test", and works on it in steps. It reads files, edits them, runs commands, reads the output, and tries again.
Think of autocomplete as a spell-checker and a coding agent as a junior teammate. The teammate is fast, but you still review every change before it ships.
Today's coding agents come in three shapes:
- Terminal agents run on your laptop inside a project folder: Claude Code, Codex CLI, Gemini CLI, Antigravity CLI and Cursor's CLI.
- Editor agents live inside an IDE: Cursor's desktop app, plus the VS Code and JetBrains extensions most vendors also ship.
- Cloud agents run on the vendor's servers and hand you a pull request: GitHub Copilot cloud agent, Codex Web and Cursor Cloud Agents.
The tools at a glance (19 November 2026)
Prices are in US dollars per month, as shown on each vendor's official pricing page on 19 November 2026. Claude and Cursor say their prices exclude tax. None of the pages showed rupee prices, so check the price at checkout.
| Tool | Vendor | Runs where | Source code | Free route | Cheapest paid plan with it |
|---|---|---|---|---|---|
| Claude Code | Anthropic | Terminal, IDE, desktop app, web | Proprietary (repo is public, "All rights reserved") | None | Claude Pro, $20 ($17 billed yearly) |
| Codex CLI | OpenAI | Terminal (also IDE, app, cloud) | Open source, Apache-2.0 | Not for the CLI (Free lists Codex only in the desktop app) | ChatGPT Plus, $20, or an API key |
| Gemini CLI | Terminal | Open source, Apache-2.0 | None since 18 June 2026 | Enterprise Code Assist licence or a paid API key | |
| Antigravity CLI | Terminal (plus app and IDE extensions) | Not stated | Individuals plan, $0, "baseline" weekly quota | Google AI Pro/Ultra (price not on the page) | |
| Copilot cloud agent | GitHub | GitHub's cloud, opens a PR | No public source found | Copilot Student, free for verified students | Copilot Pro, $10 |
| Cursor | Anysphere | Desktop editor, CLI, cloud | No public source found | Free, "limited usage", own models only | Plus, $8 (no Cloud Agents); Premium, $30 |
All six tools can use MCP servers. MCP (Model Context Protocol) is a standard plug that lets an agent use outside tools, like a database or a ticket tracker, without custom code for each agent.
Three changes in 2026 that older articles miss
1. Gemini CLI no longer has a free tier
At launch on 25 June 2025, Google gave Gemini CLI users 60 requests a minute and 1,000 a day free with a personal Google account. That tier is gone.
Google's 19 May 2026 developer blog post said Gemini CLI would stop serving free and Google AI Pro/Ultra users on 18 June 2026. Those users moved to Antigravity CLI (command agy). The Gemini CLI README now says it is for "Enterprise Code Assist only", plus paid API keys.
2. Copilot's coding agent is now "Copilot cloud agent"
GitHub announced its coding agent on 19 May 2025. The docs now call it Copilot cloud agent, and the pricing page markets it as "Copilot in the cloud". Copilot Free does not include it, but the free Copilot Student plan does.
3. Cursor is sold inside "The X Subscription"
Cursor's pricing page says it "is sold as part of The X Subscription, one subscription for Grok, Cursor, and X". The usage pool is shared with Grok and refreshes weekly. The $8 Plus plan runs Cursor's own models only, with no Cloud Agents and no third-party models.
How to install each one
These are the official commands from each vendor's docs. Read any script before you pipe it into your shell.
| Tool | macOS / Linux / WSL | Windows / other |
|---|---|---|
| Claude Code | curl -fsSL https://claude.ai/install.sh | bash | irm https://claude.ai/install.ps1 | iex (npm install is deprecated) |
| Codex CLI | curl -fsSL https://chatgpt.com/codex/install.sh | sh | npm install -g @openai/codex |
| Gemini CLI | npm install -g @google/gemini-cli | Needs Node.js 20+ |
| Antigravity CLI | curl -fsSL https://antigravity.google/cli/install.sh | bash | irm https://antigravity.google/cli/install.ps1 | iex |
| Copilot cloud agent | Nothing to install. Assign a GitHub issue to Copilot, or start a task from GitHub.com, the IDE or the GitHub CLI. | |
| Cursor | App from cursor.com/download; CLI: curl https://cursor.com/install -fsS | bash | irm 'https://cursor.com/install?win32=true' | iex |
Project instruction files, and why AGENTS.md matters
Every agent reads a plain Markdown file from your repo before it starts. You use it to say how to build, test and format the project. Think of it as the note you leave for a new teammate on day one.
| Tool | Its own file | Reads AGENTS.md? |
|---|---|---|
| Claude Code | CLAUDE.md | Yes, in place of CLAUDE.md |
| Codex CLI | AGENTS.md (32 KiB cap by default) | Yes, it is the native file |
| Gemini CLI | GEMINI.md (name configurable) | Listed as a supporter on agents.md |
| Copilot cloud agent | .github/copilot-instructions.md | Yes, the nearest AGENTS.md wins |
| Cursor | .cursor/rules (.mdc files) | Yes, as a simple alternative |
AGENTS.md is an open format that the agents.md site says is "used by over 60k open-source projects". It is now stewarded by the Agentic AI Foundation under the Linux Foundation. If your team uses more than one agent, write one AGENTS.md and you are covered almost everywhere.
Permissions and sandboxes: what the agent is allowed to do
An agent that can run shell commands can also delete files or leak a secret. Each tool has a dial between "ask me before everything" and "do anything". A sandbox is a fenced-off area: commands inside it cannot touch files or the network outside the fence.
| Tool | Default | Most permissive mode | Sandbox |
|---|---|---|---|
| Claude Code | default: reads freely, asks before edits and commands | bypassPermissions (docs: "isolated containers and VMs only") | Optional OS sandbox for Bash on macOS, Linux, WSL2; no network by default |
| Codex CLI | workspace-write: edits inside the project | danger-full-access | Built in: Seatbelt on macOS, bubblewrap on Linux, native on Windows |
| Gemini CLI | default: asks for approval | yolo, only via a flag | Opt-in (--sandbox): Seatbelt, Docker/Podman and others |
| Copilot cloud agent | Pushes only to its own branch; workflows wait for your approval | n/a (runs in GitHub's cloud) | Cloud sandbox with a firewall allowlist |
| Cursor | Sensitive actions need manual approval | "Run Everything" | Sandboxed shell in Auto-review mode, where possible |
Rule of thumb: never use the "do anything" mode on your own laptop with your real SSH keys and cloud credentials on it. Use it only inside a throwaway container or VM, which is what the Claude Code docs also say.
Which one should you start with?
- You are a student with no budget: apply for Copilot Student (free for verified students, includes the cloud agent). Also try the Antigravity CLI Individuals plan ($0) for a terminal agent. Its free quota is not published as a number, so expect limits.
- You want to learn how agents work inside: read Codex CLI or Gemini CLI. Both are Apache-2.0 on GitHub, so you can see exactly how they build prompts, call tools and sandbox commands.
- You already pay for ChatGPT Plus or Claude Pro: Codex CLI comes with Plus, and Claude Code comes with Pro. Both are $20 a month, so try the one your subscription already covers.
- You prefer an editor over a terminal: Cursor's Free plan has limited agent use. Plus is $8, but it has no cloud agents and no third-party models.
Whatever you pick, the skill that transfers is the same: break a task into small steps, write a good AGENTS.md, and review every diff.
Honest limits
- Prices and plans move fast. In 2026 alone, a free tier closed, a product was renamed, and a pricing model moved into a bundle. Check each pricing page before you pay; this page was checked on 19 November 2026.
- "Unlimited" rarely is. Claude plans share a limit that resets every 5 hours, with weekly caps. Cursor and Antigravity use weekly quotas. Long agent sessions can burn through them in an afternoon.
- Agents are confidently wrong. They will invent an API, delete a test to make it "pass", or edit the wrong file. A green test run is not a review.
- Prompt injection is real. A web page, an issue comment or a README can hide instructions that the agent then follows. Keep secrets out of reach and keep approvals on for anything that touches the network.
- Your code leaves your machine. Every tool here sends your code to a hosted model. Check your company's or college's policy before you point an agent at private code.
Frequently asked questions
Is Claude Code free?
No. Anthropic's pricing page says Claude Code is included in all paid Claude plans, and the Free plan does not list it. The cheapest is Claude Pro at $20 a month, or $17 a month billed yearly, before tax (checked 19 November 2026).
Is Gemini CLI still free?
No. Google ended free and Google AI Pro/Ultra access to Gemini CLI on 18 June 2026 and moved those users to Antigravity CLI. Gemini CLI now serves Enterprise Code Assist licences and paid API keys. The old 1,000 requests a day free tier is history.
Which AI coding agent is free for students?
GitHub Copilot Student is free for verified students and includes Copilot cloud agent, with automatic model selection only. Antigravity CLI also has a $0 Individuals plan with a weekly quota. Cursor's Free plan gives limited agent use with its own models.
Is Codex CLI open source?
Yes. OpenAI's Codex CLI is on GitHub at openai/codex under the Apache-2.0 licence. The Codex IDE extension and Codex cloud are not open source. OpenAI says the repo takes issues but not outside pull requests.
What is AGENTS.md?
AGENTS.md is a plain Markdown file in your repo that tells coding agents how to build, test and style the project. It is an open format stewarded by the Agentic AI Foundation under the Linux Foundation. Codex, Copilot, Cursor and Gemini CLI support it, and Claude Code can read it.
Is it safe to let a coding agent run commands?
Only with limits. Keep the default approval mode on your own machine, turn on the sandbox where the tool has one, and keep secrets out of the project folder. Use the 'do anything' modes only inside a throwaway container or VM.
Crio AI Engineering
Build agents that work outside a demo
Crio's AI Engineering program is project-first: you build LLM apps, RAG pipelines and tool-using agents, then learn to evaluate, secure and ship them. Talk to the team about whether it fits you.
Sources
- Claude Code overview · Anthropic
- Claude Code permission modes · Anthropic
- Claude Code memory (CLAUDE.md, AGENTS.md) · Anthropic
- Claude pricing · Anthropic
- Claude Code licence · Anthropic (GitHub)
- openai/codex repository and README · OpenAI (GitHub)
- Codex pricing · OpenAI
- Codex sandboxing · OpenAI
- Codex AGENTS.md · OpenAI
- google-gemini/gemini-cli repository and README · Google (GitHub)
- An important update: transitioning Gemini CLI to Antigravity CLI (19 May 2026) · Google Developers Blog
- Introducing Gemini CLI (25 June 2025) · Google
- Gemini CLI sandbox · Google
- Antigravity pricing · Google
- About Copilot cloud agent · GitHub Docs
- GitHub Copilot plans · GitHub Docs
- GitHub Copilot: meet the new coding agent (19 May 2025) · GitHub Blog
- Copilot repository custom instructions · GitHub Docs
- Cursor pricing · Anysphere
- Cursor rules · Anysphere
- Cursor run modes · Anysphere
- AGENTS.md · Agentic AI Foundation
Every price, date, licence and feature on this page comes from the vendor pages above, fetched on 19 November 2026. Prices are in US dollars as displayed; Claude and Cursor state they exclude tax. The 'which one should you start with' advice is Crio's own opinion.
Last updated: