How to Choose an AI Coding Agent

Choose an AI coding agent by four things, in this order: where it runs (terminal, editor or cloud), how it bills (plan limits or tokens), which models it lets you use, and how much control you keep over edits and permissions. Then run two real tasks on two finalists. Prices and plans below were checked on 31 October 2026.

What you will have at the end

• A shortlist of two tools, picked from a table instead of a hunch
• A monthly budget based on how each tool meters usage
• A seven-day trial plan with the same tasks on each tool
• A rule for when to switch, and how to switch without losing your setup

We keep deeper pages for each tool. This guide is the routing page, so follow the links when a row matches you. Prices have their own page in AI coding agent pricing compared, and the full directory is AI coding agents.

Step 1: Pick the shape

Coding agents come in four shapes, and the shape matters more than the brand.

Two changes from the past year affect the table. Windsurf is now Devin Desktop; windsurf.com redirects to devin.ai. And Roo Code shut down on 15 May 2026, with the repository archived. Zoo Code is a fork under Apache-2.0, per its README, so new setups should look at that instead.

Builders for non-developers, such as Lovable, Bolt and Replit, are a different category. If you want to ship an app without reading code, see our best vibe coding tools instead.

A simple test for shape

Ask where you want the work to appear. If the answer is "in my files, as a diff I review", an editor fits. If it is "I describe the task and come back", a terminal or cloud agent fits. If it is "I want to pick the model myself", look at the open-source group.

Step 2: Understand how each one bills

The price on the page is only half the story. The meter decides what you actually get.

Setting ANTHROPIC_API_KEY in your environment makes Claude Code bill the API instead of your plan, which surprises people who expected the plan to apply.

Entry and heavy prices

Team seats differ again. Copilot Business is $19 and Enterprise $39. Cursor Teams run $30, $100 or $200 per seat. Claude Team is $20 or $100 per seat on annual billing. Our per-tool pages go deeper: Claude Code pricing, Cursor pricing, Copilot pricing, Codex pricing and Kiro pricing.

A caution on entry plans. Cursor Plus at $8 covers Cursor's own models only, with no third-party models, while Premium at $30 adds Claude and GPT models at API price from a prepaid credit. Check what the cheap tier excludes before you compare it with a tool that includes it.

Step 3: Check the models you get

If model choice matters to you, check three things on each plan: which models are included, whether you can bring your own API key, and whether a given model is metered at API price.

• Single-vendor tools use their maker's models. Claude Code uses Claude models, Codex uses OpenAI's and Gemini CLI uses Google's. You get tight integration and one bill, and you give up model choice.
• Multi-model editors such as Cursor and Copilot offer several vendors. Third-party models may draw from a separate allowance or be billed at list price.
• Open-source tools such as Aider, Cline, OpenCode and goose let you point at an API key or a local model. You pay the provider directly, and you can use open-weight models for privacy or cost. See our open-source LLMs for coding page.

Model line-ups change often, so confirm the current list on the vendor page before you buy. Our Claude models and GPT models pages track the main ones.

Step 4: Weigh control and safety

Every agent can edit files and run commands, so look at what each lets you restrict.

Claude Code, for example, supports permission rules and hooks, and a hook's deny holds even in bypass mode. See our hooks guide. For data handling, read each vendor's terms. One example from our research: Devin's parent company says it may train on data by default on some plans, with paid plans able to opt out and Enterprise never without written consent. Read the current terms before sending private code anywhere.

If your code cannot leave the machine, look at tools that run against a local model, and at our local coding agent with Ollama page.

Step 5: Use shared project instructions

Pick tools that read a shared instructions file. AGENTS.md is read by many agents, including Codex, Cursor, Gemini CLI and Aider, per the agents.md site, and Claude Code reads it when no CLAUDE.md is present. A shared file keeps you from locking your project notes into one tool. See our guides on AGENTS.md and CLAUDE.md.

Step 6: Match the tool to the person

This is editorial judgment from the documented facts above, not a ranking from a benchmark.

One limit is worth knowing before you pick Kiro or any subscription tool. Kiro's FAQ says a subscription works with Kiro IDE, Kiro CLI, Kiro Web, Kiro Crew, ACP compatible IDEs and automation such as reviews in CI, and that use through third-party harnesses that route requests outside Kiro's own interfaces is not permitted. Read each vendor's terms for your plan before you wire a subscription into another tool.

Step 7: Run a one-week trial

Reading about agents only goes so far. Run the same work through two finalists.

1. Pick two real tasks. One small bug fix and one multi-file change from your own backlog.
2. Write the same prompt for each tool, and save it. Use the same repo and branch point.
3. Record four things per run: did the tests pass, how many corrections you had to make, how long it took, and how much of the plan limit it used.
4. Check the diff yourself. Passing tests is necessary and not sufficient.
5. Try a limit. Do a long session on day six to see how the plan behaves when it runs low.
6. Decide on day seven by cost per accepted change, not by feel.

Our Cursor vs Claude Code and Claude Code vs Codex pages cover two common finalist pairs.

A worked budget, labelled as an example

Say a solo developer wants one editor and one terminal agent. Take the published entry prices from the table. Cursor Plus is $8 and Claude Pro is $17 on annual billing or $20 monthly. That is $25 to $28 a month for the pair. If the Cursor Plus limit to its own models proves too tight, Premium at $30 raises the total to $47 to $50. If Claude's shared limits run out, the next step is Max at $100.

The point of the example is the shape of the decision. You are not choosing a tool for the year. You are choosing the cheapest pair that does not block your work this month, and you can step up in a day.

Now the same developer on a team of ten. Copilot Business is $19 a seat, so ten seats are $190 a month, before any extras. Cursor Teams start at $30 a seat, so ten seats are $300. Those are list prices from the pages above, and your negotiated or annual price may differ.

Five mistakes people make

1. Comparing headline prices. An $8 plan and a $20 plan do not buy the same thing. Compare what each includes and how it meters.
2. Judging on a toy task. Every agent can write a to-do app. Test on your own code.
3. Ignoring the permission model. A fast agent that cannot be restricted is a risk on a real repo.
4. Locking project notes into one tool. Keep the shared parts in AGENTS.md.
5. Never reading the data terms. Know where your code goes and whether it can be used for training before you paste in private work.

When to switch

Switch when one of these happens twice in a month:

• You hit the plan limit before you finish the task, on a plan you already upgraded.
• The tool keeps editing files you told it not to touch, and you cannot enforce the rule.
• A model you need is not offered, or is billed at a rate that breaks your budget.
• Your team standardizes on something else.

Moving is cheaper than it used to be if your instructions live in AGENTS.md, your tools come in over MCP, and your rules are plain Markdown. See our MCP install guide for the portable part.

Our take

Pick by shape first and brand last. Two tools with the same shape are close to each other in daily use, while two with different shapes feel nothing alike. Most developers we would expect to end up with an editor for typing and a terminal agent for the larger jobs, with one shared instructions file between them.

Do not buy a heavy tier on day one. The entry plans exist so you can learn how a tool meters before you commit $100 or more. Move up only after you have hit a real limit on real work.

Last, beware of choosing on a benchmark chart. Published scores are vendor-reported unless a third party ran them, and none of them measures your repo.

Checklist

• Chose the shape: terminal, editor, extension or cloud
• Read the meter for each finalist
• Confirmed which models the plan includes
• Checked permission rules, hooks or sandbox support
• Put shared instructions in AGENTS.md
• Ran the same two tasks on two tools
• Picked on cost per accepted change

Related pages

• AI coding agents
• AI coding agent pricing compared
• Cursor alternatives
• Free AI agents
• Write a good AGENTS.md