Claude Artifacts: Create, Test, and Share Safely

Direct answer

Claude Artifacts is the name of a Claude feature that turns a conversation into a self-contained document, code file, website, diagram, or interactive tool. A dedicated window displays the result beside the chat. You can revise it through follow-up prompts, inspect its code, switch versions, download it, and publish or share an approved version. Treat an artifact as user-generated content and a prototype until you have tested its behavior, accessibility, security, data handling, and production limits.

Summary

Purpose: move substantial reusable output out of the chat stream and into a dedicated working surface.
Formats: documents, code, single-page HTML, SVG, diagrams, flowcharts, and interactive React components.
Workflow: define the outcome, generate a small first version, inspect code, test it, revise one concern at a time, and select the exact version.
Consumer plans: publishing creates a public link that anyone can open; AI features require a Claude account and use the viewer's limits.
Organization plans: sharing stays inside the Team or Enterprise organization and can expose source-conversation attachments to authorized viewers.
Boundary: use Artifacts for exploration and validation; move durable products into a normal repository, deployment, and review process.

What is a Claude Artifact?

Claude creates an artifact when its response is substantial, self-contained, and likely to be edited, reused, or referenced later. Current documentation says this content is typically longer than 15 lines and can stand without the surrounding conversation. Claude displays it in a dedicated window to the right of the main chat.

Documents: a reusable brief, policy, lesson, plan, specification, or other Markdown or plain-text output.
Code: a focused snippet or file that benefits from syntax highlighting, versioning, inspection, and download.
Single-page websites: HTML, CSS, and JavaScript assembled into a previewable concept.
SVG images: vector output that can be inspected and copied as code.
Diagrams and flowcharts: structured explanations where relationships matter more than prose alone.
Interactive React components: calculators, quizzes, dashboards, simulators, and other stateful interfaces.

An artifact is not the same as every downloadable file Claude creates. Claude's code-execution and file-creation capability can also generate Excel, PowerPoint, Word, and PDF files in a sandbox. Artifacts are the dedicated interactive or editable surface described in this guide.

Check Anthropic's current Artifact definition, formats, editing controls, and plan availability.

Enable and find Claude Artifacts

Current Anthropic documentation requires Code execution and file creation to be enabled. Free, Pro, and Max users manage it under Settings and Capabilities. Owners of Team and Enterprise organizations control it under Organization settings and Capabilities. If an artifact does not appear, first confirm that capability before changing the prompt repeatedly.

The Artifacts section in the Claude sidebar can organize creations and start a new one. An artifact created inside a conversation does not automatically appear there. Publishing adds it to that section under the current Help Center rules. Before publishing merely to organize work, review the audience and data effects described later in this guide.

Start with an artifact brief

A prompt should define the product decision, not only the visual style. Give Claude enough structure to build a testable first version without inventing the audience, data, states, or success criteria.

Outcome: what the artifact helps a person understand, decide, calculate, practice, or complete.
Audience: who uses it, their context, and the knowledge or device assumptions that matter.
Inputs and outputs: accepted data, validation, units, defaults, empty states, and the expected result.
Content: exact facts, labels, examples, source links, and claims that need verification.
Interaction: primary path, alternative paths, errors, loading, reset, keyboard behavior, and feedback.
Constraints: privacy boundary, external dependencies, storage, connector access, responsive sizes, and browser target.
Quality bar: accessibility, readability, performance, tests, edge cases, and conditions that mean the prototype failed.

For example, ask for a responsive subscription calculator for a small product team. Define three plan inputs plus monthly and annual results. Require missing-value validation, keyboard access, clear focus states, plain-language errors, and no network or persistent storage. Ask Claude to show assumptions beside the result and include a short test checklist.

Build the smallest useful version

Ask Claude to restate the outcome, users, assumptions, states, and acceptance criteria before generating the artifact.
Create one primary workflow first. Keep secondary screens, AI features, connectors, storage, animation, and branding out of the initial test unless essential.
Open the artifact and try the main path. Use realistic input, an empty value, invalid data, a boundary value, and a reset.
Inspect the underlying code. Look for hidden network calls, hardcoded secrets, unsafe HTML, unnecessary dependencies, inaccessible controls, and fabricated content.
Request one focused revision at a time. Name the observed problem, expected behavior, affected state, and acceptance test.
Use the version selector to compare changes and return to the last known-good version when a revision causes a regression.
Download or copy the selected version only after the content, behavior, source record, and ownership are clear.

For Markdown artifacts, Anthropic documents in-place editing through highlighted text and Edit with Claude. Erroring artifacts can expose a Try fixing with Claude action that carries error details into a new message. An automated repair is a hypothesis, not proof; rerun the failing path and the wider test set after every fix.

Verify the artifact as a real interface

A convincing preview can still be wrong. Evaluate the output at the level its audience will experience, then inspect the implementation behind it.

Task success: can a first-time user finish the intended job without explanation from the creator?
Content: verify every material fact, calculation, label, unit, link, legal statement, and generated example.
States: cover empty, loading, success, partial, invalid, unavailable, permission-denied, and recovery paths.
Accessibility: use semantic controls, visible focus, keyboard operation, meaningful names, sufficient contrast, and text alternatives.
Responsive behavior: try narrow mobile, tablet, laptop, zoom, long labels, large text, and reduced-motion settings.
Security: inspect input handling, links, downloads, embedded content, storage, network calls, connectors, and copied third-party code.
Maintainability: confirm where source lives, who owns it, how it is tested, and whether a future editor can understand it.

Do not use an interactive artifact as a substitute for an editorial image. A diagram should remain a diagram with semantic text and a suitable visual asset. A calculator or prototype is appropriate only when interaction helps the reader complete the actual task.

Publishing and sharing are different

Check the account type before clicking Publish or Share. The same-looking action can create a public consumer link or an organization-only business link. Select the exact artifact version first.

Free, Pro, and Max publishing: makes the artifact public to anyone with the link. Basic interaction can work without a Claude account.
Organization-only sharing: limits Team and Enterprise access to authenticated members. Project artifacts also require access to the source project.
Embeds: a published consumer artifact can provide embed code. The publisher must list the domains allowed to embed it.
Build on a public artifact: the current flow copies its code into a new chat. The old Remix button is no longer the documented path.
Unpublish: revokes the consumer artifact and prevents that same artifact from being published again. A later publication needs a new artifact.
Unshare: removes organization access through the Artifact shared dialog without making the item public.

Organization sharing can expose more than the visible artifact. Current sharing documentation says viewers also gain access to attachments and files in the conversation that created it. Review the full source conversation, project access, uploaded material, and resulting version before sharing.

Read Anthropic's current publishing, organization sharing, embedding, copy, and unpublish rules.

Treat public artifacts as untrusted content

A public artifact is hosted on Claude infrastructure, but its content and code come from a user. In July 2026, Huntress investigated a malvertising campaign that sent searchers to a malicious public Claude Artifact. It imitated a Claude Desktop download page, redirected visitors to an attacker-controlled site, and led to SectopRAT infections at at least 29 organizations. Anthropic removed the reported artifact.

Identity: a `claude.ai` artifact URL identifies the hosting surface, not the trustworthiness or affiliation of its creator.
Downloads: navigate to a vendor's official website yourself. Do not install software supplied or linked by a public artifact.
Links: inspect the destination and purpose before leaving the artifact, especially for login, payment, extension, and download prompts.
Code: copy only from a trusted creator, inspect every dependency and capability, and run it in a disposable restricted environment.
Data: avoid secrets, credentials, personal data, customer files, or internal material unless the exact storage and sharing boundary is approved.
Reporting: use Anthropic's content-reporting path for a suspicious artifact and notify your security team after any attempted download or execution.

Review Huntress's technical account of the July 2026 FakeAgent artifact campaign.

Understand AI-powered artifacts

An AI-powered artifact can call a limited text-based Claude capability. Anthropic's AI documentation states that Claude writes the code and the app runs on its infrastructure. Users authenticate with their own Claude accounts. Artifact publishers do not provide an API key or pay for other people's interaction. Each viewer's use counts against that viewer's Claude limits.

This model is useful for a prototype, but it does not remove product obligations. Define what the AI may answer and what data it receives. Specify error behavior, outputs that require verification, and the experience when a user has no remaining usage. Test harmful input, prompt injection, unsupported requests, ambiguity, and recovery.

Follow Anthropic's current AI-powered Artifact tutorial and prototype-to-production boundary.

Review MCP access separately

On supported paid plans, an artifact can use MCP to read from or write to connected services on web and desktop. Anthropic documents a first-use approval prompt and says each user authenticates each MCP server independently. Organization admins can enable or disable Artifact MCP access, but current documentation says they cannot choose the specific servers an artifact may use.

Confirm the connector publisher, server URL, requested scopes, data destinations, retention, and revocation path.
Use a least-privilege test account and records created for evaluation rather than a production calendar, inbox, task system, or workspace.
Require confirmation before writing, sending, deleting, inviting, changing access, or acting on externally supplied instructions.
Revoke authorization after the trial and verify that the connector no longer reads or changes data.

Choose personal or shared storage deliberately

Persistent storage supports stateful trackers, journals, and collaborative tools on supported paid plans. Personal storage separates each person's data. Shared storage exposes the same data to every user of the artifact and triggers a confirmation on first interaction. Each data path uses the mode chosen by the creator.

Current documentation sets a 20 MB limit per artifact and accepts text only, not images, files, or binary data. Storage does not work during development; the artifact must be published. Unpublishing permanently deletes its personal and shared storage. Use disposable data to verify the published storage boundary before asking real users to depend on it.

Know when to leave the Artifact prototype

Anthropic's own tutorial positions Artifacts as strong for prototyping and demonstration. A production service usually needs a normal engineering and operations boundary that one conversational artifact cannot supply by itself.

Source control: durable history, code review, ownership, repeatable builds, and dependency updates.
Identity: explicit authentication, authorization, roles, tenant isolation, session control, and account recovery.
Data: documented schema, migrations, encryption, backups, deletion, retention, privacy review, and regional requirements.
Reliability: monitoring, rate limits, retries, timeouts, fallbacks, incident response, and service-level expectations.
Security: threat modeling, secret management, dependency scanning, content security policy, audit logs, and abuse controls.
Quality: automated tests, accessibility checks, browser coverage, performance budgets, staged deployment, and rollback.
Product operations: support, analytics, consent, moderation, legal terms, billing, and a named owner for future changes.

Copy the approved code into a repository, replace prototype assumptions with production services, and preserve the validated user problem rather than the generated implementation. Rebuild the tests around the intended behavior. Then review, deploy to a non-production environment, and run the real user flow before release.

Claude Code Artifacts are a separate workflow

Claude Code Artifacts are available on Team and Enterprise plans. A Claude Code session can publish a live private page that updates as the session continues. Examples include a pull-request walkthrough, a dashboard from session data, or an investigation timeline. By default, only its creator can see it. The creator may share it inside the organization, but Anthropic says it cannot be shared publicly.

Compare Claude Code web, desktop, terminal, and mobile surfaces in our interface guide.

Troubleshoot Claude Artifacts

No artifact appears: enable Code execution and file creation, confirm organization policy, and request a substantial self-contained output explicitly.
Artifact is buried in chat: open it from the conversation. Publishing, not creation alone, adds a consumer artifact to the sidebar library.
A change broke the preview: use the version selector, return to the last working version, and ask for a narrow fix with a reproducible test.
Persistent storage fails: current storage works only after publishing and only within the documented plan, size, and text-data limits.
MCP access is missing: confirm a supported plan and surface, organization access, the user's connector authentication, and the first-use approval.
Team member cannot open a shared item: check organization membership, authentication, artifact share state, and source-project access.
Public artifact seems official: read its user-generated disclaimer, verify the creator independently, and avoid its installer, login, or payment links.

Frequently asked questions about Claude Artifacts

Core Artifact access and Claude-powered artifacts are documented across Free, Pro, Max, Team, and Enterprise plans. MCP, persistent storage, Claude Code Artifacts, publishing, and organization sharing have plan-specific rules. Check the current Help Center for the exact capability you need.

On Free, Pro, and Max plans, publishing creates a link anyone can view and interact with. Business-plan artifacts use organization-only sharing on Team and Enterprise. Live Cowork artifacts have their own organization-only rules.

The hosting domain does not verify the creator's claims or code. Treat a public artifact as untrusted user-generated content. Avoid downloads, credentials, sensitive data, copied code, and external links until you verify the publisher and inspect the behavior.

It can validate a single-page concept or interactive tool. A durable public product usually needs source control, production hosting, identity, data governance, monitoring, tests, accessibility, security review, support, and controlled deployment.

Sources and verification

Claude Help Center, What are artifacts and how do I use them?: definition, formats, enablement, editing, AI, MCP, storage, and Claude Code behavior.
Claude Help Center, Publish and share artifacts: public and organization audiences, embeds, copying code, attachments, unpublish, and unshare rules.
Claude Help Center, Create and edit files with Claude: separate file-generation capability, network controls, sandbox, and prompt-injection guidance.
Anthropic tutorial, Prototype AI-Powered Apps with Claude artifacts: AI setup, iteration, user usage, and prototype-to-production boundary.
Huntress, Inside FakeAgent, July 2026: observed public Artifact malvertising path, fake installer, affected organizations, takedown, and malware attribution.
Current guides from Zapier, UsingClaude, and AI Toolbox: competitor structure, examples, reader questions, and workflow framing; not proof of Anthropic product claims.

Connect Better Design