Claude Code

The choice comes down to two profiles: see which one is yours.

Who it is for. Claude Code is built for work on a live project - at work or your own. Redo half a site, clean up tangled code, fix what broke: the agent finds the right place, plans, and checks itself. It handles long tasks without supervision, reporting along the way.

Don't write code? The terminal is optional: you can assign a task in the desktop app or directly in the browser, like in a chat. But the product is still about code - to accept the agent's work, you need to understand what it did.

Who it is not for. There are no inline suggestions as you type here; for those, people go to AI editors. There is no free Claude Code: at Anthropic only the chat is free, the agent starts at $17/mo. Claude models only, you can't plug in your own local ones.

Just getting started? Haven't written code or built projects yet - start with the page“What is vibecoding”, then come back here for a tool.

It is a fit. Install it following section 3. If you are choosing between it and Codex, see the decision guide in 7.

Rewrote the page in a day. The Codex page was a month old, and its facts were outdated. The agent reread 128 of our news stories about it, checked every number against fresh sources, and rewrote the page from scratch - on August 17, it was current again.

Delivered research overnight. It received a question in the evening and delivered a report by morning: about forty search queries, 45 sources read, and a conclusion added to the page. It worked on a separate laptop while people slept.

Compiled a biography and verified the facts itself. Founder biography in English: I cross-checked academic papers against the arXiv database and deal dates against reporting from those years. Anything I could not verify, I said so.

Found where the money was leaking. The hosting bill was melting before our eyes: $15 out of $20 in six days. The agent checked billing, found the cause, and suggested another plan - $103 less per month.

Fixed what failed overnight. A nightly demand-measurement script started failing. The agent found the network cause, rewrote the script with retries - the failures stopped.

On duty without days off. Claude Code also runs the site's news feed: on a separate laptop, it checks every minute whether a story is ready, selects stories, and writes copy without a human.

What it looks like. The loop in any of these tasks is the same: you write what you need, the agent lays out a plan, makes edits, shows every change line by line, and sees it through. The full loop is in the frames below.

How much freedom to give it. By default, the agent asks for confirmation for every edit and command. Once you trust it, turn on auto mode: edits are accepted automatically, stopping only for risky actions.

Where it stops. The agent does not take dangerous actions even in auto mode: in one of our runs, it refused to click the permissions-granting button - a human made the final click.

Then on its own. Further up the axis: loops, background tasks and parallel subagents in one session. The top step: cloud sessions at claude.ai/code keep running when the laptop is closed, and you can steer them from your phone. Since Sep 17, Pro and Max have a beta of Projects, where Claude itself hands out parts of one job to parallel cloud sessions.

If a game project needs 3D objects, pick a separate model generator. Export terms and rights are covered in the Tripo card.

Where the code goes. Under commercial terms (Team, Enterprise, API), Anthropic does not train models on your code. On personal Pro and Max plans, training happens only with consent: data is retained for 5 years when enabled, 30 days when disabled.

What the agent can break. It edits files and runs commands. Since August 14, new sessions run in auto mode: a separate classifier filters out dangerous commands. You set the boundaries yourself - down to confirming every step; the enterprise plan adds organization-wide controls.