What is a Claude artifact?

What Claude puts in an artifact

Claude reaches for an artifact when the output is substantial, standalone, and something you are likely to edit or reuse rather than read once. In practice that means:

• Documents. Reports, specifications, long-form writing, anything with structure you will come back to.
• Web pages. A single file of HTML, CSS and JavaScript that renders live in the panel.
• React components. Interactive interfaces that actually run rather than being described.
• Diagrams and charts. SVG, flowcharts, sequence diagrams, data visualisations.
• Small tools. Calculators, dashboards, explainers, games. Anything self-contained enough to work without a server behind it.

The unifying idea is one self-contained file. An artifact is something that reads or runs on its own, with no build step and nothing to install. That constraint is why artifacts appear so fast, and it is also the source of most of their limits.

Why it is a panel and not a message

The panel is not a display decision, it is a versioning one. Chat is append-only: ask for a change and you get a new message, leaving you to work out which of the six blocks of code on screen is the current one. An artifact replaces itself. Ask for a darker header and the header goes dark in the thing you are already looking at.

That single behaviour is what makes artifacts feel like work rather than conversation, and it sets an expectation that matters later. Once you have felt a thing update in place, going back to emailing report-v3-FINAL.html to somebody is genuinely painful.

Where an artifact actually lives

This is the part most explanations skip, and everything practical follows from it.

An artifact lives on Anthropic’s servers, attached to your account and to the conversation that produced it. It is not a file on your computer. It is not a website you own. It is a view into Claude, and reaching it normally means reaching your Claude account.

For the thing artifacts are designed for — building something with Claude and looking at it — that is exactly right, and the speed comes from it. The friction starts when the artifact stops being something you are looking at and becomes something you need to hand over.

Who can open one

By default, only you, inside the conversation. Publishing changes that: Claude mints a claude.ai link and anyone holding it can open the artifact in an ordinary browser, with no Claude account needed.

What is missing is the middle. There is no per-person access, no passcode, and no way to restrict a link to one company’s email domain. The choice is private to you, or open to anyone who ends up with the URL — including anyone your recipient forwards it to. We have written up whether published artifacts are public and what to do when that is not the access model you wanted.

The four limits worth knowing

None of these are complaints about artifacts. They are consequences of the design, and they only bite in one situation: when the artifact is a deliverable rather than a demo.

• The address is not yours. A published artifact sits on claude.ai. For internal work that is fine. When the recipient is a client, the domain is part of what you are delivering.
• Access is all or nothing. Private to you, or open to anyone with the link. No passcode, no named recipients, no company-domain restriction.
• Unpublishing is one-way. Per Anthropic’s help centre, once you unpublish, the Publish button deactivates and that version cannot be restored; if the artifact used persistent storage, that data is permanently deleted. We cover the unpublished-and-cannot-republish message separately.
• It depends on somebody else’s uptime. When you see artifact storage is temporarily unavailable, nothing you did caused it and waiting is usually the fix. But if that artifact was the thing due this afternoon, the outage is still your problem.

When an artifact should become a page

The honest rule: leave it as an artifact while you are building and showing. Move it when you are delivering.

Four signals that the moment has arrived. It needs to sit on your own domain. It must reach one named person and nobody else. It has to outlive the conversation and the account that made it. Or it will be revised, and whoever you sent it to must keep seeing the current version at the address you already gave them.

Plain disclosure, since this is our own site: Stacktree does that last part. You publish the artifact’s HTML and get a private, unguessable URL that opens with no account for the reader, can carry a passcode or a company-email gate, can sit on your own domain, and can be replaced in place so a link you sent last week shows this week’s version. An agent can do it directly through MCP, or you can move one by hand in about thirty seconds — the walkthrough is at how to share a Claude artifact.

If none of those four signals apply to what you are making, publish it from Claude and get on with your day. That is what the button is for.