Artifacts in Claude Code, and the sharing modes that exist between you and the whole internet.

What artifacts in Claude Code are

Artifacts in Claude Code let the agent turn what just happened in a session into a live, interactive web page. Anthropic's examples are dashboards, incident timelines, PR walkthroughs, and checklists, the kinds of things you would otherwise paste into a doc or a channel. The page is built, in Anthropic's words, using the full context of your session, including your codebase, your connectors, and the conversation itself. So it is not a generic export; it is shaped by what the agent already knows about your work.

Two details make it feel alive rather than static. First, updates land in place: when Claude Code updates an artifact, the open page refreshes, so a teammate watching a long-running task sees it change without reloading. Second, each artifact keeps version history you can restore from, and all of them sit in a gallery. The result is closer to a living view of an agent's work than a one-shot file.

It runs from the Claude Code CLI and the desktop app, and the pages themselves are viewable in any browser, subject to the access rules below. One mechanical detail matters more than it looks: Claude writes each artifact as an HTML or Markdown file in your project, then publishes it. The source of every artifact is already a file in your repo.

One boundary worth stating, because it sets expectations: an artifact is a page, not an app. It is one self-contained page served under a strict Content Security Policy: no scripts, styles, or images from other hosts, no fetch or WebSocket calls, no backend, no stored form input, no multiple routes, and a 16 MiB ceiling on the rendered page. The single exception is MCP connector calls, which the page hands to claude.ai to make on its behalf, so a page can show live data without making network requests itself. In other words it is a self-contained static page, which is exactly the shape that travels well if you ever need to host it somewhere else.

How sharing and access work

This is the part that changed on 13 July 2026, when public sharing shipped, so it is worth reading in the docs' current words. Three rules define the model:

• Private by default. "A new artifact is visible only to you." Sharing is a deliberate act from the page header, which names you as the author.
• Two audiences. On Team and Enterprise, grant access to specific people or everyone in your organization; viewers sign in to claude.ai as org members, and you can promote a viewer to editor, which is the multiplayer half of the July launch. Or share a public link "that anyone on the internet can open, with no claude.ai sign-in required."
• Plan asymmetry. On Pro and Max, "a public link is the only way to share an artifact": no org audiences, no editors. On Team and Enterprise it runs the other way: public sharing is off until an Owner enables external sharing for the organization, and turning it back off blocks access through existing public links.

One further boundary: an artifact that pulls live data through MCP connectors cannot be shared to a public link on any plan. On Pro and Max, that means a connector-backed artifact stays private to you.

Notice what is absent from the model. The audience options are you, your organization, or everyone; there is no gated mode between the last two. No password, no expiry, no link that one client can open without the whole internet being able to. Until public sharing shipped, Anthropic's docs answered the outside-the-org question with a sentence worth preserving: "ask Claude for the HTML file and share that file directly." The docs have since replaced it with the public link, but the advice still stands for every case between org-only and fully public, and the mechanics make it trivial, because the artifact already is an HTML file in your project.

What it is great at

Taken on its own terms, this is a strong answer to a real problem: the output of an agent session is usually trapped in a terminal or a transcript, and getting it in front of teammates means copying, pasting, and reformatting. Artifacts skip that.

• Session-aware, not generic. Because the page is built from your codebase, connectors, and the conversation, it reflects the actual work rather than a blank template.
• Live, not a snapshot. The refresh-in-place behaviour means a long task can be watched as it unfolds, which is exactly what an incident timeline or a build dashboard wants.
• Versioned. History and a gallery mean an artifact is durable and recoverable, not a throwaway, and the Share control can pin viewers to a version or always show the latest.
• Access controls that fit internal work. Org audiences, editor roles, admin toggles, role scoping, and retention policies match how companies actually gate internal tools, and the public link covers the demo case.

For a Team or Enterprise org that lives in Claude Code and wants to surface an agent's work to colleagues, this is a clean fit, and an open primitive like ours does not replace it.

What is still missing: the gated middle, domains

With public sharing shipped, the questions cluster in a new place: not "can anyone outside see it" but "can exactly one person outside see it."

Can you share an artifact with someone outside your org without making it public? No. The documented audiences are your org (Team and Enterprise, sign-in required) or anyone on the internet. A client under NDA, a contractor, a candidate: each of them sits in the gap between those two. The route through the gap is the one Anthropic's own docs used to prescribe: take the artifact's HTML file, which already exists in your project, and host it on a private link with a gate the recipient passes without an account. On Team and Enterprise plans where an Owner has not enabled external sharing, it is also the only route to any outside viewer at all.

Can you put it on your own domain? Custom domains are still not mentioned anywhere in the docs. Artifacts are hosted by Anthropic at claude.ai URLs, and shared pages carry Claude's viewer chrome with your name and gallery link. The honest reading remains "not offered," not "never."

Can you export it or self-host it? This one now has a structural answer: the docs state that Claude writes the page to an HTML or Markdown file in your project before publishing. There is no export button because none is needed; the source file is already yours, sitting in your repo, and it hosts anywhere.

None of this is a criticism of the feature for its stated purposes, internal work and public demos. The gap is the deliverable in between: the page that has to reach someone specific, gated, on a domain that is not Anthropic's. If that is your situation, the docs do not cover it, and the HTML file is your way through.

Three "Claude artifacts" features, untangled

It is easy to conflate three different things Anthropic now ships under the word "artifacts." They are not the same:

• Claude.ai artifacts are created in a chat on claude.ai. On Free, Pro, and Max they publish to public links; on Team and Enterprise they cannot be published publicly at all, and unpublishing is permanent. If you have been sharing Claude artifacts, this is usually the one you mean.
• Live Artifacts live in Cowork and are persistent dashboards that re-query your connected apps and refresh when you open them. We cover those in what are Claude Live Artifacts.
• Artifacts in Claude Code, this post, are built from a coding session; private to you, then shared to org audiences (Team and Enterprise) or a public link, with no gated mode in between.

The sharing rules differ across the three, so the one that matters for "can I send this to someone outside the company?" is which surface you made it on. The full grid is in the sharing matrix.

Where this leaves open tools like Stacktree

Full disclosure: we build Stacktree, so treat this as us describing where we sit, not neutral advice. Stacktree solves a narrower, more open problem than Artifacts in Claude Code, and the two address different needs.

Stacktree is a publish primitive for agent-made HTML, exposed over MCP, so it works across Claude Code, Cursor, Codex, and Claude.ai rather than inside one product. Its sharing model fills the middle Claude's two audiences skip: a page is private by default, the unguessable URL is itself the credential, and a viewer needs no account and no org membership. You can add a password, an email-domain gate with magic-link verification, an expiry, or end-to-end encryption; map a custom domain on the Pro plan; replace a page in place with update_site; and self-host the whole stack on your own Cloudflare account, since it is source-available. The trade is scope: Stacktree hosts static HTML only, with no session-aware live data layer. So if you want a live, session-aware internal page for your Claude org, or a public demo link, Artifacts in Claude Code is built for that; if the page needs to reach one gated recipient, carry your domain, or stay portable, an open primitive is the closer fit. They are not substitutes.

What to watch

This reflects Anthropic's docs as of 18 July 2026, and this feature moves fast: it has already widened twice, from a Team and Enterprise beta to all paid plans in early July, then to public-link sharing on 13 July. What remains worth rechecking:

1. A gated external mode. Public sharing answered "can anyone outside see it." The most-asked question is now a password, an expiry, or a single-recipient link; none exist today.
2. Custom domain support. Still unmentioned in the docs, and it is the top roadmap request in the launch replies.
3. Connector-backed pages beyond the org. Today an artifact that calls MCP connectors cannot be shared publicly on any plan; if that changes, live dashboards start traveling.
4. Editors beyond Team and Enterprise. Pro and Max public links are view-only; multiplayer editing is currently an org-plan feature.
5. Use beyond Claude Code. Today it is built from a Claude Code session (or a Claude Tag session); if the surface widens, the picture changes again.

If any of these land, the answers above will need updating. For now, this is the documented state of the feature.