AI Coding Agents: The Directory and How to Choose

An AI coding agent reads your code, edits files and runs commands to finish a task you describe, with you approving the risky steps. The big choices are price model, open or closed source, which models it can use, and where it runs. This guide compares nine, with every fact checked on 31 October 2026.

The nine, side by side

More agents are on their own pages: Windsurf, Aider, Cline, Roo Code, Goose, Devin, Amp and OpenHands. App builders such as Lovable, Bolt and Replit Agent are a different category, covered in vibe coding.

What counts as an agent

A chat assistant answers questions. A completion tool finishes the line you are typing. An agent does a loop: it reads files, decides what to change, edits them, runs a command such as your tests, looks at the result and tries again. The Claude Code docs describe the idea as a tool that "reads your codebase, edits files, runs commands, and integrates with your development tools."

Most products here do more than one of those. GitHub Copilot still gives inline suggestions and also has an agent mode. Cursor is an editor with an agent and cloud agents. The word "agent" in a product name tells you less than the surfaces listed in the table.

OpenClaw is the odd one. It is an assistant that lives in chat apps, and coding is one of the things it can drive. We include it because people searching for open-source agents land there.

Choose by what matters to you

Price

The cheapest paid doors are not the same size.

Look at what the number measures, not the number. Copilot and Kiro sell credits. Claude and Codex sell usage windows. Cursor sells a weekly pool. Antigravity sells rate limits. They cannot be compared dollar for dollar, which is why we keep a pricing page for each: Claude pricing, Copilot pricing, Kiro pricing.

License and source

The open ones let you read the code, pin versions and fork. The closed ones are often more polished, and in two cases (Claude Code and Antigravity) the public repository or download is not the source. Claude Code has a public GitHub repo whose license says all rights are reserved.

Open source matters most when you work in a regulated company, need to audit what runs, or want to host the whole stack. It matters least if you just want the best result this week. See open source AI agents.

Models

If you want to change model vendors freely, that narrows the list to Copilot, Cursor, Kiro, OpenCode and OpenClaw. If you are loyal to one lab, its own agent is the natural first stop. For a deeper model comparison see compare models.

Where it runs

• Terminal first: Claude Code, Codex CLI, Gemini CLI, OpenCode, Cursor's agent and Kiro's kiro-cli.
• Editor first: Cursor, GitHub Copilot, Antigravity IDE, Kiro IDE.
• Cloud agents: Cursor and Copilot both run work in the cloud; Claude Code has web sessions and scheduled cloud tasks; Codex has Codex Web.
• Chat apps: OpenClaw.

MCP, skills and instruction files

MCP is the common language for add-ons. Every agent on this page supports it in some form, with different commands.

AGENTS.md is the closest thing to a shared standard: several of these read it. If you write one good file, you can move between agents with little work. Start your add-on search with the MCP server list and the best MCP servers, then try Playwright MCP, GitHub MCP and Context7.

Privacy and data

Every hosted agent sends your prompts and the code it reads to a model provider. What differs is retention and training terms, which depend on the plan.

A rule that fits all of them: do not paste secrets, and read the data terms for your exact plan before using work code.

The price ladder at three usage levels

Entry prices hide where each plan ends. These are the monthly plan ladders from the pricing pages our pricing writers checked on 31 October 2026. Read them as steps, not as equivalents, because each vendor measures usage differently.

The pattern is that the first rung is cheap and the third is about ten times the second to last. Heavy agent use, where the tool runs for hours on large tasks, is what pushes people up the ladder. If you only ask for small edits and explanations, the first or second rung is usually enough. If you leave an agent running on multi-file work all day, price the top rung before you start, or you will hit a limit mid-task.

Two traps are worth naming. Credit plans such as Copilot and Kiro charge more for heavier models, so the same credits last a very different time depending on the model you pick. Usage-window plans such as Claude and Codex reset on a schedule, so a limit hit on Tuesday afternoon can mean waiting. Neither is better. Choose the shape that matches how you work.

What to put in AGENTS.md

AGENTS.md is the one file most of these agents read, so it is where a shared setup starts. Keep it short and put only what the agent cannot work out from the code. A starting point:

# Commands
- Install: npm install
- Test: npm test (run before saying a task is done)
- Lint: npm run lint

# Conventions
- TypeScript strict mode; no `any`
- Components in src/components, one per file
- Never edit files under generated/

# Rules
- Ask before adding a dependency
- Do not touch the database migrations folder

Three habits keep this file useful. Write commands exactly as you would type them, because the agent will run them. Add a line when an agent makes the same mistake twice, and remove lines that no longer matter. And keep one copy: if Claude Code needs CLAUDE.md, point it at the shared file instead of duplicating the content, so the two never drift. The table above shows which agents also want their own file, such as GEMINI.md or Cursor rules.

Moving from one agent to another

Switching is cheaper than it looks, because the valuable parts are mostly portable.

A sensible way to trial a second agent is to run both on the same branch for a week, with one change at a time and a human reviewing every diff. Cancel the loser when you are sure, not on the first bad afternoon. Keep your decisions in version control, in AGENTS.md, a short design note or the commit messages, and the next switch costs almost nothing.

Rolling an agent out to a team

An agent on one laptop is a personal tool. On a team, it needs rules.

1. Pick a pilot group. Three or four developers on one service for two weeks, with the same ticket types, so you can compare.
2. Write the permissions down. Decide which commands may run without asking, which may not, and which files are off limits. Every agent here has a permission model, and the defaults are rarely right for a production repository.
3. Decide where code may go. Name the plans and vendors that your company's data rules allow, and ban pasting secrets. A personal consumer account is not the same as a business plan, and the data terms differ.
4. Keep review human. Agent-written code goes through the same pull request review as any other code, and the author owns it.
5. Measure something real. Time to merge, review rework and escaped bugs say more than how people feel. Track them before and after.
6. Set a budget owner. Usage-based plans can surprise you, so someone should watch spend monthly.

Read AI coding agent pricing compared when you reach step 6, and the individual agent pages for each vendor's security and privacy documents.

Terms you will meet

The last row explains much of this page. Two agents can use the same model and still behave differently, because the harness decides which tools the model sees and how it recovers from errors. That is why we compare the agents, not only the models.

Which one should you try first

Head-to-head pages

• Cursor vs Claude Code
• GitHub Copilot vs Claude Code
• Claude Code vs Codex
• Claude Code vs OpenCode
• Gemini CLI vs Claude Code
• Antigravity vs Claude Code
• Kiro vs Claude Code
• OpenClaw vs Claude Code
• Cursor alternatives

How to test an agent fairly

A comparison is only useful if the test is yours. A simple method:

1. Pick one real task from your backlog, small enough to finish in an hour.
2. Write it as a ticket with acceptance criteria, and give every agent the same text.
3. Use the same model where you can. Many agents let you choose Claude or GPT, and holding the model fixed isolates the agent.
4. Compare the diff, not the chat. Count how many corrections you gave.
5. Record cost: credits or usage windows consumed.
6. Repeat on a second, different kind of task, such as a bug fix and a refactor.

Our pages are built from vendor documentation and pricing, not from this kind of test, so treat them as a way to shortlist two or three, then run the test yourself.

Our take

This is an editorial reading of vendor pages, not a benchmark.

The biggest shift in the last year is convergence. Terminal agents grew IDE extensions, editors grew CLIs and cloud agents, and almost everyone reads AGENTS.md and speaks MCP. That lowers the cost of picking wrong, because moving is cheap. It also means the differences that last are the dull ones: how usage is billed, which models you can reach, what the vendor does with your data, and whether you can read the code.

So choose on those four, shortlist two, and run your own one-hour test. If you want a single default, match the agent to the subscription you already have: Claude plan, ChatGPT plan, GitHub seat or a Google licence. If you have none, begin with a free tier.

Checklist

• Decide your must-haves: open source, model choice, price model, privacy
• Shortlist two agents from the tables above
• Write one shared AGENTS.md for your repository
• Add one or two MCP servers, not a dozen
• Run the same task in both and compare the diffs
• Re-check prices on the vendor pages the day you subscribe