Claude artifacts now share publicly and edit multiplayer.

What Claude actually shipped

The announcement from @ClaudeDevs is one line: "Artifacts now support public sharing and multiplayer editing in Claude Code, and can be created with Claude Tag." In the launch thread, Thariq shows the use case that matters: a project dashboard created as a Claude Tag artifact, then edited by teammates or by a local Claude Code session against the same artifact. Credit where it is due, this is a good feature and people want it. The replies are full of "finally," and one is worth quoting because it is the whole point: "public sharing without team/enterprise is awesome." Sharing a Claude artifact used to require a team plan; now it does not.

It is also the seventh major entrant in twelve months to converge on the same premise, that AI makes HTML and the HTML needs a URL: Shopify's internal Quick, OpenAI's Codex Sites, here.now, Vercel Drop, Notion's HTML block, Cloudflare Drop, and now Anthropic itself. When the model vendor ships hosting for its own output, the category is settled. The question is no longer whether AI output needs a URL. It is which of these URLs you can actually hand to a client.

Since launch day, Anthropic's Claude Code documentation has caught up with the announcement and now states the sharing model precisely. A new artifact is private to its author. On Pro and Max, in the docs' words, "a public link is the only way to share an artifact." On Team and Enterprise, sharing defaults to people inside the organization, editors included, and public links stay off "until an Owner enables it for the organization." One further boundary: an artifact that calls MCP connectors at view time cannot be shared to a public link on any plan. That gives us enough certainty to draw the whole matrix.

The sharing matrix, surface by surface

Four surfaces now carry the name "Claude artifacts," each with different sharing rules, and the differences are exactly where people get caught. This is the verified state as of 18 July 2026, from Anthropic's docs and help center:

Read the matrix by columns and the pattern is stark. Every Claude row has "None" or "org authentication" in the gates column and "No" in the domain column. The rows differ only in which flavor of ungated audience you get: yourself, your org, or everyone. The gated middle, one recipient, no seat, your domain, is the whole fourth column of the last row, and it is the shape of a client deliverable.

It shares publicly. A deliverable cannot be public.

The launch word is "public sharing," and the now-published docs confirm the flow has no password, no email-domain gate, and no expiry. That is fine for a demo you want the timeline to see. It is the wrong default for the work most people are actually publishing. Agent output routinely embeds real data: a client's numbers, an internal dashboard, a prototype under NDA. The moment the link is public, obscurity is the only protection, and obscurity is not a gate.

This is Stacktree's first default and it runs the other way. Every published URL is unguessable and private by default. When a page should reach one recipient and no one else, you add a password or an email-domain gate the recipient passes without a Claude account, or an expiry that revokes it after the engagement. A public link cannot be un-published into a private one; a private-by-default link can always be made public. The direction of the default is the whole difference between a demo and a deliverable.

It lives on claude.ai, not your domain.

A shared Claude artifact carries Claude's domain and, for free users elsewhere in the product, Claude's branding. In the launch thread, the top roadmap question is exactly this: a user asking Anthropic whether custom domains are coming, "the interesting version is artifacts served under our own domain (CNAME + TLS)." As of launch, they are not there.

For an agency handing a report to a client, the domain is the product. The deliverable has to read as the agency's work, on the agency's domain, with no third party's badge on it. Stacktree serves the same HTML on your own custom domain with the badge removed, so the page communicates "this came from us," not "this was generated in a chat." That is the single most-requested thing in Anthropic's own launch replies, and it is a property of the host, not the model.

It is Claude-only.

The feature shares Claude-authored artifacts from Claude Code. That is the right scope for Anthropic and the wrong scope for anyone whose stack is not all-Claude. The same week, OpenAI put Sites in Codex and ChatGPT, Cursor keeps its own agents, and most real workflows mix them. Output from Codex, Cursor, or any MCP agent does not get Claude's sharing, and it never will, because a model vendor ships hosting for its own model.

A publish primitive is deliberately agent-agnostic. Stacktree takes the HTML any agent produces and puts it on a URL over MCP, whether Claude, Codex, Cursor, or a custom loop wrote it. As each vendor builds native sharing for its own output, the cross-agent layer is the one position none of them will occupy, because occupying it means hosting a competitor's output.

Multiplayer editing vs an agent-regenerated artifact

Multiplayer editing, several people in one artifact at once, is a genuinely good addition, and Claude solving it for humans is worth using. But note the shape it solves. As one reply in the thread put it, "multiplayer for humans is solved, multiplayer for agents isn't." The high-value artifact is often not one a room of people co-edits. It is one an agent regenerates: a report rebuilt nightly, a dashboard the agent updates as data lands, a status page that changes without a human touching it.

That shape wants replace-in-place, thirty revisions at one stable URL rather than thirty links, and it wants the agent to publish the update itself over an API or MCP rather than a person clicking share. Multiplayer human editing and agent-driven regeneration are different needs. Claude just shipped the first; the second is still the publish-primitive half of the category.

Which fits the thing you are holding

The boundary is the same one that sorted every prior entrant. If the thing you are holding is a Claude artifact you want teammates to see and co-edit, in a chat, publicly, Claude's new sharing is the natural home and you should use it. It removed the team-plan wall and it is good at that job.

If the thing you are holding is a deliverable, a report for a client, a dashboard your agent regenerates, a prototype that carries real data and needs a gate rather than an audience, the four properties flip: private by default, your own domain, any agent not just Claude, and a publish the agent can call. That half is what Stacktree is built for, and Anthropic shipping public sharing makes the line sharper, not softer.