What are Claude Artifacts? Create, edit, and share them

A Claude Artifact is a substantial, self-contained output that opens in its own panel beside your conversation. Instead of burying a document, code file, diagram, or interactive tool inside the chat, Claude gives you a separate result you can inspect and revise while the conversation continues.

That is the simple definition. The complication is that Anthropic now uses artifact for several related experiences with different rules. A normal artifact in Claude, an AI-powered artifact, a Claude Code artifact, and a Cowork live artifact are not interchangeable names for the same thing.

The rest of this guide explains the ordinary artifact workflow first, then shows where those other meanings diverge. Product behavior and availability were checked against Anthropic's documentation on August 19, 2026.

What counts as a Claude Artifact?

Anthropic's current artifact guide describes an artifact as content that is substantial, self-contained, useful outside the immediate answer, and likely to be edited or reused. It says the content is typically longer than 15 lines, but length is a signal rather than a format rule.

The supported examples are broad:

• Markdown and plain-text documents;
• code snippets;
• single-page HTML;
• SVG images;
• diagrams and flowcharts; and
• interactive React components.

That last item creates an easy misunderstanding. An interactive React artifact is not automatically an AI-powered artifact. A calculator whose JavaScript runs in the browser can be interactive without calling Claude. It becomes AI-powered when the artifact invokes Claude to generate or interpret something for its user.

A short answer, a conversational explanation, or a quick list usually does not need the separate panel. If you do want a reusable result, ask directly:

Create this as a Claude Artifact. Keep the result self-contained so I can
revise, inspect, and share it separately from our conversation.

Then name the real deliverable: a decision memo, an SVG process map, an interactive pricing calculator, or a single-page prototype. “Make an artifact” chooses the workspace; it does not supply the brief.

Create, edit, and keep the right version

Artifacts require Code execution and file creation to be enabled in Claude's capability settings. If Claude keeps replying in the chat instead of opening the side panel, check that setting before rewriting the prompt repeatedly.

1. Step 1CreateAsk for a self-contained artifact in a conversation, or start one from the Artifacts area.
2. Step 2ReviseDescribe the change in chat; for Markdown, select exact text and use Edit with Claude for a targeted request.
3. Step 3InspectUse the version selector, choose among multiple artifacts in the chat, and check the underlying content or code.
4. Step 4Export or distributeCopy or download the result, or select the intended version before you Publish or Share it.

Revise it without starting over

Ask Claude to update the artifact in the same conversation. Changes appear in the artifact window, and the version selector lets you return to earlier revisions. If one conversation contains several artifacts, its controls let you choose which one Claude should reference or change.

For Markdown documents, you can also highlight a passage and use Edit with Claude. That is useful when “tighten this paragraph” is clearer than describing its location in the chat. Anthropic notes that changing the artifact does not rewrite Claude's memory of the earlier content, so tell Claude when a revised fact should also guide the rest of the conversation.

Inspect, copy, or download it

The artifact window exposes its underlying content or code. You can copy the result, copy its code, or download it for work outside Claude. If an artifact throws an error, Try fixing with Claude sends the error back into the chat, but it is a debugging attempt, not a guarantee that the next version will work.

The old Remix action for somebody else's published artifact is no longer the current workflow. Anthropic's publishing guide now tells users to copy the code into a new chat. That creates an independent artifact; it does not edit the original.

Publishing also changes where you find it

An artifact created in a conversation does not automatically appear in the Artifacts collection in your sidebar. Publishing it adds it there. Until then, keep the source conversation if you expect to return to the draft.

This is why “I made it” and “I can find it from the Artifacts area” are separate states. Choose the correct version before publishing merely to organize it, because Publish also changes who can access it on personal plans.

The four artifact meanings in more detail

1. A regular artifact is a separate output

This is the core feature: Claude places a substantial result beside the chat so the conversation remains the instruction layer and the artifact remains the thing being made. Documents, code, diagrams, HTML, and React components all fit that model.

Publishing or sharing is optional. You can keep revising in the source conversation, copy the content elsewhere, or download the result. When you do distribute it, the regular-artifact rules in the next section apply.

2. An AI-powered artifact can call Claude

An AI-powered artifact is a subtype of the regular artifact. Claude writes the tool, Anthropic runs it, and people who use its AI features authenticate with their own Claude accounts. The creator does not paste an API key into the artifact, and each user's requests count against that user's Claude limits.

AI power is different from ordinary interactivity. A static quiz can check an answer with browser code; an AI-powered tutor can generate a new explanation for the learner. That distinction matters when you tell recipients whether they will need to sign in.

Some published artifacts can also use persistent storage and connections to outside services:

• Storage is available only after the artifact is published. Anthropic documents a 20 MB limit per artifact, text-only data, and separate personal and shared stores.
• A personal store keeps each user's data separate. A shared store makes data visible to the artifact's users and shows a confirmation the first time.
• Connections use the Model Context Protocol (MCP). Each user authenticates the connected service independently, even when the artifact itself is shared.

If that protocol is unfamiliar, this plain-English MCP guide explains why the connection belongs to the user rather than being a universal password inside the artifact.

3. A Claude Code artifact is live session output

Claude Code uses the same name for a different publishing surface. According to Anthropic's current Claude Code artifact documentation, it creates a self-contained .html, .htm, or .md file in the project and publishes that session output to a Claude URL. The page can update at the same URL as the session continues, and each publication becomes a version.

The page is useful for annotated pull-request walkthroughs, investigation timelines, implementation comparisons, and dashboards built from session data. It is not a hosted application: Anthropic documents one self-contained page, no backend, no stored form submissions, no multiple routes, strict external request limits, and a maximum rendered size of 16 MiB.

A later Claude Code session needs the artifact URL to update the existing page; without it, the session creates another artifact. Connector-backed Claude Code artifacts also have narrower sharing because each viewer's connector permissions control the live data.

4. A Cowork live artifact is local and persistent

A live artifact in Cowork is an interactive HTML view built around ongoing work: a tracker, dashboard, comparison tool, or reference that can refresh from connected services. Anthropic's Cowork live-artifact guide says it is a paid-plan feature in Claude Desktop and lives locally on the computer that created it. It does not follow you to another device.

Its sharing rules are separate again. Pro and Max live artifacts cannot be shared or published; Team and Enterprise can share them inside the organization. A recipient opens the result in Claude Desktop and uses their own access to any connected services.

The word live here means a persistent working view that can refresh. It does not mean a public website.

Publish and Share are different access states

For normal artifacts in Claude and Claude Desktop, Anthropic currently uses two different actions by plan:

• Publish on Free, Pro, and Max creates a public link. Anyone with the link can use basic functionality without signing up. AI-powered capabilities require a Claude account and use the viewer's limits.
• Share on Team and Enterprise keeps access inside the organization. Viewers sign in with the organization account, and an artifact created in a Project also requires access to that Project.

A published artifact can be embedded on another website, but the publisher must list which domains are allowed to embed it. An embed is another view of the same artifact; it does not move ownership or hosting to that website.

Before you distribute an artifact, check these five things:

1. Version: open the exact version you intend to distribute before choosing Publish or Share.
2. Source material: when an organization artifact is shared, viewers also gain access to attachments and files from the conversation that created it.
3. Viewer identity: basic public interaction can be anonymous, but AI calls, organization access, and connectors require the appropriate sign-in.
4. Stored data: know whether the artifact uses personal or shared storage before entering sensitive information.
5. Connections: every viewer authorizes connected services for themselves; your access does not silently become theirs.

Unsharing an organization artifact is a different action. Do not transfer the irreversible Publish rule to every surface, and do not assume a Cowork or Claude Code artifact follows the regular-artifact access model.

Artifact vs. Project vs. ChatGPT Canvas

These concepts overlap because all three can surround a substantial piece of work. Their primary jobs are different.

• Claude Artifact: the output. It is the document, code, visual, or interactive result you are iterating on and may distribute.
• Claude Project: the context workspace. Anthropic describes Projects as workspaces with chat histories, project knowledge, uploaded context, and instructions. A Project can contain chats that create artifacts; it is not itself the artifact.
• ChatGPT Canvas: an editing interface in a different product. OpenAI's current Canvas guide describes a workspace for writing and code that supports direct edits, highlighted requests, inline feedback, versions, previews, sharing, and downloads. The useful distinction here is not which product wins: Canvas is the editing surface, while Claude uses Artifact as a family of output and publishing experiences.

Choose a Project when several chats should share background and instructions. Choose an Artifact when one result deserves its own revision surface. If you use ChatGPT instead, Canvas is the comparable place to edit a substantial result, but its feature and sharing rules remain OpenAI's, not Anthropic's.

When the artifact is enough

An artifact is often the finished deliverable. Keep it in Claude when you need a reviewable document, an explorable visualization, a temporary prototype, a session walkthrough, or a small tool whose Claude access model is acceptable to everyone using it. Moving it elsewhere would add work without improving the job.

Reconsider when people depend on an address you control, the result needs multiple pages, or its lifecycle should not be tied to the artifact surface. That is a separate ownership decision, and the artifact-link-versus-website guide owns the full comparison. Use the same-chat hosting guide for the exact transfer and publication steps when the source is a compatible static Artifact in a Claude conversation you control. Those steps are not part of defining an Artifact.

One nuance matters: do not apply Claude Code's “no backend” rule to every AI-powered artifact. Regular published artifacts can use Claude, connected services, and Anthropic's bounded text storage. Conversely, do not treat those capabilities as proof that a Claude Code session page or a Birta static site is a general full-stack application.

Frequently asked questions