Claude Code is Anthropic's coding agent. It works inside your project: it reads files, edits them, runs commands and tests, and shows you what changed. You use it from the terminal, an editor, the desktop app or the web. It asks your permission for actions depending on how you set it up.
A question from the app. Pick an answer.
What are an agent's permissions for?
Fewer questions can speed things up, but that is not their job. Permissions set limits.
For example, it may read and edit, but not go online. Which ones it asks you about first is up to approvals.
The model is a separate setting. Permissions are about actions: read, edit, run.
You ask "fix the bug in the sign up form". A chat without access to your project gives you code to paste. Claude Code has tools to work on the files: it finds the right one, edits it, runs the tests and tells you the result. It is a coding agent: a language model with tools to read, write and run, which decides the next step by looking at what happened.
It suits multi step tasks: fixing a bug, adding a feature, writing tests, tidying code or explaining a project you do not know. For a quick question, a chat is enough.
There are several doors to the same agent: the terminal, extensions for editors such as VS Code and the JetBrains ones, the desktop app and the web, which is handy for leaving a long task running. Version-controlled rules travel with the repo; local settings do not carry over to the cloud on their own.
Which ones exist today and how to install them changes often: check the official Claude Code documentation before you start.
The project rules (conventions, commands, what not to touch) go in a CLAUDE.md file it reads at the start. Permissions set which actions it can take on its own and which it asks you about first, like editing a file or running a command. Give it the minimum the task needs.
For big changes, ask for a plan first: in plan mode it explores the project and prepares a proposal before implementing, and you correct it. At the end, review the file diff and the command results.
Access depends on where you use it: a supported Claude subscription or, where allowed, pay per use with the Anthropic API or other providers. Which fits depends on how much you use it and whether it is yours or a team's; the full comparison is in subscription or API. Plans and prices change: check the official page.
Codex, Cursor and GitHub Copilot also work as agents in your project, each with different model options, integrations and pricing. The core idea is the same: read, edit, run and show you the result. What changes is the model behind it, where you use it, how comfortable it feels and what it costs for your case.
To choose, try two on a task of your own, not a demo, and compare the diff and the time it took you to review it. Rules files can be shared too: see AGENTS.md.
"If it says it is done, it is done." "Done" is a claim: ask for the tests to be run and read the diff. "More permissions, less friction." Until it deletes what it should not: grant the minimum. "A huge CLAUDE.md makes it better." It loads in full every session and dilutes what matters: keep it short. "It works equally well on everything." A big, vague task goes better split into steps, with a plan first.