Coding CLIs: the terminal as the interface

A chat window gives you an answer. You copy it into your editor, run it, something fails, you copy the error back. That loop is most of the time you spend, and you are the one carrying the messages.

A coding CLI closes that loop. It runs in your project folder, so it can open your files, edit them, run your tests, read the failure, and go again. You stop being the courier. That is the whole pitch, and it is also the whole risk: something that can edit files and run commands without asking you each time is powerful in exactly the way that is hard to undo.

What they all have in common

They differ in which model they use and what they cost, and the directory covers that per tool. Underneath, they are the same shape: They work in a folder. Run one in your project and that project is its world. They ask before doing something irreversible, at least by default. Read the thing it is asking about. The prompt is the safety feature. They read an instructions file. This is the part people skip, and it is the part that decides whether the tool is any good at your project.

Where each one keeps its things

Two separate concerns: its own settings and sign-in, which are yours, and the instructions it reads about this project, which belong to the team. Pick your tool.

The instructions file is the whole game

A terminal agent with no instructions re-derives your project from scratch every session: how to run the tests, where things live, what the conventions are. It will guess, and it will guess the industry default rather than your actual setup. Ten lines fixes most of it. The build and test commands. Where the important code lives. Two or three rules it keeps getting wrong.

Try one properly

The first session people run is usually "write me a feature", which is the hardest thing to judge. Start with something where you already know the right answer, so you can tell whether it did well.

The third one is the one to actually keep. Have it draft the file, read it yourself, delete half of it, and commit what is left.

Picking one

Honest answer: start with whichever one matches a subscription you already pay for, and switch later if it annoys you. They are close enough that the choice matters less than the habits, and your instructions file mostly moves between them. The comparisons that matter — pricing, which models, whether you can bring your own key, whether it runs local models — are on each tool's page in the directory, where they get kept up to date.

• Browse coding CLIs
• Global or project: where config lives
• AGENTS.md: a README for your coding agent
• MCP servers: giving your AI tool hands