An AI coding agent does more than autocomplete. Anthropic describes Claude Code as a tool that “reads your codebase, edits files, runs commands, and integrates with your development tools.” GitHub says Copilot can “plan and make changes, review pull requests, and automate repository tasks” (GitHub docs). Cursor calls itself “a coding agent for building ambitious software” (Cursor docs). The category is real, crowded and changing monthly, so this guide focuses on the choices that stay useful: where the agent runs, how you pay, what happens to your code, and how to keep a human in control.
How this guide was made
We did not run a hands-on benchmark and we publish no scores or pass rates. Everything below comes from the vendors' own documentation and pricing pages, read on 19 November 2026, plus the open standards linked in the text. Plans, limits and model names change often, so treat prices as a snapshot and confirm them on the vendor page before you buy. Some links in this article are affiliate links (marked as sponsored); they never change what we write. See our affiliate disclosure.
1. What an AI coding agent actually is
A chat assistant answers questions about code you paste in. An agent works inside your project: it reads files, proposes a plan, edits several files, runs your tests or build commands, looks at the output and tries again. The practical consequences are bigger than the marketing suggests. An agent can save hours on well-defined chores such as adding tests, renaming an API across a codebase or upgrading a dependency. It can also make a confident mistake across twenty files at once, so the surrounding workflow (branches, tests, review) matters more than the choice of tool.
Two open standards make agents less of a lock-in decision than they used to be. The Model Context Protocol (MCP) is “an open-source standard for connecting AI applications to external systems,” and its documentation lists Claude, ChatGPT, Visual Studio Code and Cursor among supporting clients. AGENTS.md is “a simple, open format for guiding coding agents,” a README for agents that holds build steps, test commands and conventions; its site lists Cursor, Aider, Gemini CLI and GitHub Copilot's coding agent among compatible tools, says the format emerged from work including OpenAI Codex, Cursor and Factory, and says it is now stewarded by the Agentic AI Foundation under the Linux Foundation. Writing one good AGENTS.md is therefore time well spent whichever tool you pick.
2. Four ways to run an agent
| Style | Examples | Best when | Main trade-off |
|---|---|---|---|
| IDE-centred | Cursor; GitHub Copilot in VS Code and JetBrains | You want to see diffs inline and steer the agent while you code | Tied to an editor or plugin; usage rules differ per plan |
| Terminal-native | Claude Code, OpenAI Codex CLI, Gemini CLI, Aider | You live in the shell, use any editor and want scriptable workflows | Less visual; you need to be comfortable reviewing git diffs |
| Repo or cloud agent | GitHub Copilot cloud agent, Cursor cloud agents, Claude Code on the web | You want to hand off a task and review a pull request later | Needs repository access and clear task descriptions |
| Open-source, bring your own key | Aider, Cline, Gemini CLI, Codex CLI | You want model choice, local models or no vendor seat fee | You pay model API usage directly and manage setup yourself |
3. The main options, based on official sources
Cursor
Cursor's docs describe an agent that can help you understand a codebase, plan and build features, fix bugs and review changes, with customisation through plugins, skills, MCP servers and rules, and integrations with GitHub, GitLab, Azure DevOps, Bitbucket, JetBrains, Slack and Linear. Its models table lists models from several providers (Anthropic, Meta, Moonshot and Cursor's own models among them), so you are not limited to one model family.
Pricing has changed. The Cursor pricing page we read on 19 November 2026 lists individual tiers called Free, Plus ($8/month), Premium ($30/month), Super ($100/month) and Ultra ($200/month), plus Teams from $30 per user per month and custom Enterprise. The page states that third-party models are billed separately by usage, and the individual tiers bundle other Grok features. If you remember an older “$20 Pro” plan, it was not on the page we read, which is a good reminder to check before relying on any price you saw elsewhere.
On privacy, Cursor's security page says Privacy Mode is available to free and paid users and that, when enabled, Cursor will not train on your data. It also states that Cursor holds SOC 2 Type II, ISO/IEC 27001:2022 and ISO/IEC 42001:2023 credentials, with reports available on request through its trust portal. Those are vendor statements: if compliance matters to you, request the reports rather than relying on a blog post, ours included. Check Cursor’s current plans.
Claude Code
Claude Code runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web (Anthropic docs). Anthropic recommends a native install script and notes that native installs update automatically in the background. The terminal CLI, VS Code and JetBrains also support third-party model providers. Anthropic's pricing page lists Claude Code on the paid individual plans: Pro at $17/month with an annual subscription ($200 billed up front) or $20 if billed monthly, Max at $100/month with five times Pro usage, and Ultra at $200/month with ten times Pro usage. Usage limits apply and, as Anthropic notes, how quickly you hit them depends on the model, the length of your chats and the size of your files. If you prefer to pay per token, the API route exists too. See Claude plans.
GitHub Copilot
GitHub's plans documentation lists Copilot Free, Student, Pro ($10/month), Pro+ ($39/month) and Max ($100/month) for individuals, and Business ($19 per seat) and Enterprise ($39 per seat) for organisations. It states that all plans include Copilot CLI and the Copilot app, and that plans come with an allowance of “GitHub AI Credits,” with usage-based billing beyond it. The cloud experience lets one session move between GitHub.com, mobile, the Copilot app, VS Code and the CLI, and can run in a cloud sandbox, a GitHub Actions runner or a custom environment. GitHub notes that from 28 October 2026 the new cloud experience is available to paid plans, while Free users keep the older chat experience. Copilot is the natural default if your team already lives in GitHub issues and pull requests.
OpenAI Codex
OpenAI's Codex pricing page says Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans, with a usage allowance that eligible plans can extend with credits or usage-based billing. It also says sessions that authenticate with an API key are billed at API prices instead of drawing on your ChatGPT allowance. The Codex CLI is open source under the Apache-2.0 licence (per its GitHub repository). Because Codex rides on a ChatGPT subscription, it can be cost-effective if you already pay for ChatGPT, but check the allowance for your specific plan.
Open-source agents: Aider, Cline and Gemini CLI
Aider is a terminal pair-programming tool that builds a map of your codebase, commits changes to git with generated messages so you can diff or undo them, can run linters and tests after edits, and connects to cloud or local models. Cline describes itself as an “autonomous coding agent as an SDK, IDE extension, or CLI assistant,” and Gemini CLI as “an open-source AI agent that brings the power of Gemini directly into your terminal.” All three repositories are published under the Apache-2.0 licence according to GitHub's repository metadata. The tools are free to download, but you still pay for the model: an API key, a subscription the tool supports, or the hardware to run a local model. That makes costs flexible and sometimes unpredictable, so set spending limits at the provider.
4. At-a-glance comparison
| Tool | Where it runs | How you pay (per vendor page, 19 Nov 2026) | Worth knowing |
|---|---|---|---|
| Cursor | Editor app, cloud agents | Tiered plans; third-party models billed by usage | Multiple model providers; Privacy Mode; MCP and rules |
| Claude Code | Terminal, IDEs, desktop, web | Included with Claude Pro, Max and Ultra, or API billing | Third-party providers supported in CLI, VS Code, JetBrains |
| GitHub Copilot | IDEs, GitHub.com, CLI, mobile | Free to $100/month for individuals; per seat for teams | Cloud sandboxes and Actions runners; AI-credit allowances |
| OpenAI Codex | CLI, ChatGPT surfaces | Included in ChatGPT plans, credits or API billing | CLI is Apache-2.0 |
| Aider / Cline / Gemini CLI | Terminal or IDE extension | Free software; you pay your model provider | Apache-2.0; model choice, including local models |
We deliberately do not rank these tools. A fair ranking would need controlled, repeatable tasks on your own codebase, and results shift with every model release. The sections below show how to run that comparison yourself in an afternoon.
5. How to choose: a five-question checklist
- Where do you work? If you want agent edits inside your editor, start with an IDE-based option. If you prefer the shell or CI, start with a terminal agent.
- How do you want to pay? Flat plans are predictable but have usage limits. Usage-based credits and API keys scale with work done and can surprise you. Check each tool's spending controls before the first big task.
- What are your data rules? Read the privacy and retention terms for the exact plan you will use. Business and enterprise tiers often differ from individual ones.
- Which model flexibility do you need? Some tools lock you to one provider's models; others support several providers or local models. This matters if prices or quality shift.
- Who reviews the output? Whatever tool you choose, decide before you start who reads the diff and which automated checks must pass.
6. A safe workflow for agentic coding
These practices are our editorial recommendations, drawn from how the vendors' own documentation frames agent work (branches, tests, review gates and scoped permissions), not results from a test suite.
- One task, one branch. Give the agent a narrowly scoped job on its own branch so a bad run is easy to discard. Aider's automatic git commits are one example of tooling built around this idea.
- Write an AGENTS.md. Put install, build and test commands and your style rules in one file so every agent starts with the same context.
- Make tests the judge. Tell the agent which command proves the work is done, and keep a human reviewing every pull request.
- Limit permissions. Prefer read-only or approve-each-command modes for unfamiliar repositories, and keep production credentials out of the agent's environment.
- Treat MCP servers like third-party code. They can read data and take actions on your behalf. The MCP project publishes security guidance worth reading before you connect anything sensitive.
- Watch spend. Long sessions that re-read large files burn credits. Use the provider's budgets or alerts, and review usage weekly in the first month.
Security note
Autonomous tools that run shell commands and read your files carry real risk: a prompt injected through a web page, issue or dependency could try to make an agent do something you did not intend. Keep secrets out of the working directory, review commands before approving them, and prefer sandboxed or cloud environments for untrusted code.
7. Which kind of user fits which approach
- Solo developers and freelancers: an IDE-based agent or a terminal agent on a flat plan keeps budgeting simple. Try the free tiers before paying.
- Teams already on GitHub: Copilot's issue-to-pull-request flow and organisation policies are the lowest-friction place to start.
- Terminal power users and scripters: Claude Code, Codex CLI, Gemini CLI or Aider fit shell workflows and CI jobs.
- Cost-sensitive or privacy-sensitive users: open-source agents with your own API key, or a local model, give maximum control at the price of more setup.
If you are building a small business around these tools, our guide to the lean solopreneur SaaS stack shows how to budget software spend, and our guide to AI research and SEO tools covers the content side.
8. Common mistakes to avoid
- Judging a tool from one impressive demo instead of five of your own real tasks.
- Letting an agent touch authentication, payments or database migrations without a human reading every line.
- Skipping tests because “the agent already ran them.” Run them yourself in a clean environment.
- Buying an annual plan before you have used the product for a month on real work.
- Assuming a price or limit from a review (including this one) is current. Check the vendor page.
Frequently asked questions
Is Claude Code better than Cursor?
There is no honest single answer. Claude Code is a terminal-first tool from Anthropic; Cursor is an editor-based product that can use models from several providers. Which is better depends on whether you prefer the shell or an editor, how you want to pay, and which models you trust on your codebase. Try both on the same three real tasks.
Can I use Claude models inside Cursor?
Cursor's documentation lists Anthropic models among those it supports, with third-party model usage billed by usage. Availability and pricing change, so check the models table in the Cursor docs.
Are open-source coding agents really free?
The software is free under permissive licences such as Apache-2.0, but you pay for model usage through an API key or run a local model on your own hardware. Costs depend entirely on how much you use the agent.
Will these tools train on my code?
It depends on the vendor, the plan and the settings. Cursor states that Privacy Mode prevents training on your data; other vendors have their own terms. Read the data-use policy for the exact plan before using proprietary code.
What is the cheapest way to start?
Use a free tier first (several vendors offer one), or bring your own API key to an open-source agent with a low monthly spending cap. Upgrade only after a few weeks of real use.
How should I verify AI-written code?
Review the diff like any pull request, run your tests and linters, and add tests for the change. Do not merge anything you do not understand.
Sources and last-checked date
All pricing and feature statements were read from the pages below on 19 November 2026. Vendors change plans without notice.