This guide explains what an artifact is, how versions and the sidebar work, the four ways people try to export one, and where each of those routes runs out of road when the artifact is a work deliverable rather than a demo.
What is a Claude artifact?
An artifact is a rendered output panel that Claude creates alongside the conversation instead of inside it. When Claude decides the thing it is producing is a standalone piece of work — an HTML page, a React component, a Python script, a Markdown document, an SVG or a Mermaid diagram — it moves that output into a separate window on the right of the chat, where it renders live rather than sitting as a wall of code in the message thread.
Two properties make it more than a formatting trick:
- It renders. HTML artifacts display as an actual page. A diagram displays as a diagram. You are looking at output, not source.
- It is versioned. Ask for a change and Claude rewrites the artifact in place, keeping the previous version. You can step back through revisions rather than scrolling the chat for the good one.
People search for this under several names — Claude AI artifacts, Anthropic artifacts, "what is Claude artifacts" — and they all point at the same panel. Functionally, it is Claude's document canvas, the place where the deliverable lives while you iterate on it.
When Claude creates an artifact instead of inline output
Claude reaches for an artifact when the output is substantial (roughly more than a short snippet), self-contained, and likely to be reused or edited outside the chat. A three-line CSS fix stays inline. A full landing page becomes an artifact. Explanatory prose stays in the conversation; a formatted document you asked it to write becomes an artifact.
The heuristic is not perfect, and you do not have to accept it. Asking directly — "put that in an artifact" — works, as does the reverse. If you are generating something you intend to publish, ask for the artifact explicitly rather than hoping the threshold trips.
The artifacts sidebar and version history
Artifacts you create accumulate in a sidebar tied to your account, so you can reopen one from a previous conversation without hunting through chat history. Inside a single artifact, a version control lets you move between revisions — useful when the fourth rewrite is worse than the second, which is most fourth rewrites.
Both of these are session and account conveniences, not storage guarantees. If Claude cannot write to artifact storage, the panel fails to save; we cover that failure mode separately in "Artifact Storage Is Temporarily Unavailable".
Artifact types, with a worked HTML example
The types you will actually meet at work:
- HTML pages — proposals, one-page microsites, dashboards, pricing comparisons. The type that matters most for sales and marketing.
- React components — interactive prototypes and calculators, rendered live.
- Documents — Markdown or plain text: briefs, training material, meeting summaries.
- Diagrams — SVG and Mermaid: org charts, funnels, architecture sketches.
- Code — scripts and queries in any language, syntax-highlighted but not executed.
A concrete pass. You paste a CRM export into Claude and ask for a one-page account summary for a renewal call: revenue trend, open tickets, three talking points, styled to fit on a screen. Claude returns an HTML artifact. You ask for the ticket table sorted by age — version 2. You ask it to drop the logo block because the brand colors are wrong — version 3. Version 3 is the deliverable.
Now it has to leave the chat. That is where the four exits come in.
The four ways to get an artifact out of Claude
Each of these works. Each fails differently.
1. Copy and paste
Copy the artifact contents into an email, a doc, or your CMS. Fine for code you are pasting into an editor. For a rendered HTML page it is the worst option: you are moving source into a place that will not render it, and inline styles rarely survive the trip into an email client. (Related reading on why: why HTML emails hurt deliverability.)
2. Download the file
You get an .html or .md file on your desktop. It renders correctly when opened locally, which makes it feel like a solved problem. Then you attach it to an email, the recipient's gateway strips or quarantines the attachment, and you have no idea whether they opened it. An attachment is a dead end for anything you need to follow up on.
3. The built-in share toggle
Claude can publish an artifact to a shareable URL. This is the fastest route and the one that surprises people. Publishing makes the artifact accessible to anyone holding the URL — there is no recipient list, no password prompt, no expiry date, and no view data coming back to you. It is a public page with an unguessable address, which is a meaningfully different thing from a private one. We unpack that distinction in public links vs controlled access.
4. Export to a hosting layer
Take the artifact's HTML and publish it somewhere that treats it as a document with a lifecycle: an owner, an expiry, an access rule, an audit trail, and a view log. This is the route that treats an artifact as a deliverable, and it is the only one of the four that answers "who saw this."
Limits worth knowing before you send one
Four caveats catch people out, and all four matter more the more valuable the artifact is.
- Storage is tied to your account, not to a workspace. Artifacts live with the conversation and the person who made it. There is no team library, no ownership transfer when a rep leaves, and no central place where IT can see what has been produced.
- Published artifacts are public by default. No password, no domain restriction, no way to revoke access to a copy someone already loaded.
- There is no expiry. A pricing page published in March is still live in November, quoting numbers that stopped being true in Q2.
- There is no view data. You cannot tell whether the artifact was opened once, twenty times, or forwarded to a competitor.
And one more that belongs to whoever generated the input: an artifact contains whatever was in your prompt. Paste a CRM export and the customer names, emails, and phone numbers are now in the rendered page. Paste a config file and the API key is in there too. Nothing in the chat window scans for that.
How to turn a claude artifact into a durable, tracked link
The export route has two entry points, and you pick based on whether a human or a script is doing the sending.
From inside Claude, via MCP
MCP (Model Context Protocol) is the standard that lets Claude call an external tool directly. With the HTMLvault MCP server connected, Claude can publish the artifact it just built without you copying anything: scan_html checks the markup first, create_link publishes it with the settings you name, and get_analytics reports back on views later. Setup is covered in the MCP server guide.
In practice you finish iterating and say: "Publish that as a private link, expire it in 14 days, password protect it." Claude runs the scan, publishes, and hands back the URL in the chat.
From anywhere else, via REST API
If the artifact is produced by an automation rather than a person — a nightly report, a Clay enrichment run, a Zapier step, a Custom GPT action, a Gemini workflow — POST the HTML to the REST API with an API key and the same options apply. Anything that can make an HTTP request can publish an artifact.
Either way, the settings worth naming on the call are the same four:
- Scan before publish. The regex scanner checks nine categories — SSN, financial, API key, passport, address, person, DOB, email, phone — at zero token cost. It is the step that catches the CRM export you pasted three prompts ago. Teams and Enterprise can layer an AI scan on top using their own Anthropic, OpenAI, or Google key.
- Set an expiry. Anywhere from one hour to never. Match the window to the deal rather than defaulting to forever.
- Decide on a password. A pricing page that will be forwarded internally is a good candidate for password protection.
- Choose the domain. Pro includes one white-label domain; Teams gets one to three depending on seat band; Enterprise gets three. Client-facing artifacts should not arrive on a generic URL.
Published links are never indexed by search engines, and retention runs from auto-delete to two years on Pro, so the artifact's lifespan is a decision you make rather than one that defaults to permanent.
Why this matters for the person sending it
An artifact is a good rendering surface and an incomplete delivery mechanism. That is not a flaw in Claude — the panel is doing exactly what it was built to do, which is let you see the work while you shape it.
What changes when you export is who has the information afterwards. The rep who sent the proposal can see it was opened four times from two cities, and that the reader stopped scrolling at the pricing section — the kind of signal covered in per-link analytics. The marketer knows the campaign page came from their own domain. The RevOps lead knows the lead list expires on schedule instead of circulating indefinitely. And IT has an answer to the question they always ask, which is what happens to this page after it leaves the building.