AI for Writing Code: Comparing the Tools

There are many tools, and comparing them by feature lists is useless. Categories are more useful: within a category tools are similar, between categories they are fundamentally different.

The categories

Editor completion. Suggests continuations as you type. Speeds up typing without changing how you work. Useful to everyone and satisfying to almost nobody as the only tool.

Chat with a model. A separate window you carry code into. Flexible, nothing to install, all the context-moving work is yours. The model cannot see the project and does not know whether its answer compiles.

An IDE agent. An editor with an agent mode: changes appear as diffs, with code navigation and a debugger at hand. Low learning curve, high visibility. Tied to an editor session.

A terminal agent. Works on the project as a whole, runs commands, reads test output, walks the repository. The only category you can embed in a script or CI. No visual control.

A cloud agent. Takes a task, works in its own environment, returns a pull request. Good for background work; the feedback loop is slow.

What separates them in practice

Three questions divide them better than any table:

Can the tool check itself? Chat cannot. An agent can, if it may run tests. This is the main quality difference: a tool that sees test output fixes its own mistakes without you.

Does it see the whole project? Completion and chat do not. Agents do. On multi-file work the difference is decisive.

Can it be automated? Only terminal and cloud agents. You cannot schedule an editor.

How to choose

Four criteria instead of rankings:

The shape of your tasks. Local edits favour an editor. Repository-wide work and reacting to command output favour a terminal agent. Understanding an unfamiliar codebase does too.
Whether the model can be swapped. A tool locked to one provider is a risk: price and availability are outside your control.
How permissions work. A tool that runs commands unasked and offers no configuration cannot be used on a client project.
A test on your own task. Take work you have done by hand and know the right answer to, and run it through two tools. Half an hour of that beats a week of reviews.

It is also worth separating tool from model. Most tools run on several models, where the difference is reasoning depth and price, while tools differ in the way you work. More in choosing a model.

What people usually settle on

The common pairing is an editor for hands-on work plus a terminal agent for larger tasks. They do not compete; their centres of gravity differ. The detailed comparison is in Claude Code or Cursor, and the terminal category is surveyed in alternatives to Claude Code.

The one rule when running two tools: never on the same files at the same time.

On local models

A separate category by hosting rather than by tool type. You can run a model locally and point a client at it. Justified when the code cannot leave your machine.

Honestly about quality: on agentic work local models are noticeably weaker than cloud ones. An agent must hold a long chain of steps together, which is the most demanding use there is. More in local agents.

Choosing between specific options: the best AI for code. The overview is in the vibe coding guide.

FAQ

Which AI writes code best?

The question conflates two things: the tool and the model behind it, and most tools run on several models. Tools differ in how you work; models differ in reasoning depth. Choose both, by different criteria.

Are there free AI coding tools?

Free comes in three forms: provider trial allowances, editors with a limited free tier, and local models on your own hardware. Only the last is genuinely free, and it is also the weakest on hard tasks, especially agentic ones.

How does chat differ from an agent for writing code?

Chat returns text you move into the project by hand and does not know whether it works. An agent opens files, runs commands and reads output, so it can check itself. On tasks with tests the difference in outcome is large.