AI Tools · Claude Code

Your Coding Session Is Now a Link You Can Send

Anthropic just shipped a feature that quietly deletes one of the most annoying parts of working with an AI coding agent: explaining what it did. Now you just send the page.

On June 18, 2026, Anthropic added Artifacts to Claude Code. The pitch is almost too simple to sound impressive: ask Claude for one during a coding session, and it publishes your work as a single, live web page at a private link. You open it in any browser. You send the link to a teammate. They watch it update as you keep working.

That’s it. And that’s exactly why it matters.

Think about the last time an AI agent refactored a module, traced a bug, or walked a pull request for you. To share that, you had to translate it — copy snippets into Slack, screenshot a terminal, type a summary that’s already out of date by the time someone reads it. Claude Code Artifacts replace all of that with one URL that’s always current.

Why this is different from “AI made a webpage”

Plenty of tools can spit out HTML. The thing that makes a Claude Code artifact useful is where it comes from. It isn’t generated from a blank prompt — it’s assembled from three things already sitting inside your session: your actual repository, the tools you’ve connected through MCP, and the conversation that produced the work.

So a page can annotate a real diff line by line, reflect what a connected tool actually returned, and capture the reasoning that led there. It’s a record of the session, not a pretty approximation of it. When Claude produces a new iteration, the open page refreshes in place and even holds the viewer’s scroll position. Every publish is saved as a version at the same link, so a teammate watching it sees the work evolve without you re-sending anything.

The clever part is what they left out

Most of the coverage leads with “you can share a live page.” The sharper story is what Anthropic deliberately didn’t build. An artifact has no backend. A strict security policy blocks every outbound request — no external scripts, fonts, or images, and no fetch, XHR, or WebSocket calls at all. CSS and JavaScript are inlined; images are embedded directly into the page.

That seal is the feature, not a limitation they’re racing to remove. A page that can’t phone home can’t leak data or pull in a compromised script — a much smaller attack surface for something circulating inside a company. The trade-off is everything a normal app does at runtime. As Anthropic puts it plainly: an artifact is a capture of work, not an application.

Boris Cherny, who leads Claude Code at Anthropic, described how he’s been using it: “visual explanations of tricky code, system diagrams, quick previews of a few animation options, data analyses and dashboards I share with the team.” The common thread is finished work meant to be read once and acted on — not operated.

When to use one — and when to build a real app

This is the question the feature actually raises, so here’s the honest line between the two:

The deciding factor is almost always the same: does anything need to persist, authenticate beyond your organization, or call an API when someone opens it? If yes, you need a deploy. If it’s finished work people read and act on, it’s an artifact. That’s also the cleanest way to understand how it differs from OpenAI’s Codex Sites, which launched a couple of weeks earlier and bet the opposite way — persistent backends, databases, public sign-in. One is a review canvas; the other is a place to host a product. Most teams will end up using both.

The catch nobody puts on the box

A page that looks authoritative still has to earn it. An agent-generated dashboard can be confidently wrong, and a clean render makes the mistake easier to miss, not harder. Artifacts make it effortless to circulate an agent’s work — which means the discipline of actually reviewing that work has to scale right alongside it. Treat a shared artifact as a prompt for review, not a settled answer, and it’s a genuine upgrade. Treat the polish as proof, and you’ve just scaled the exact problem it looks like it solves.

Two more constraints worth knowing before you plan a rollout: it’s Team and Enterprise only at beta (no Free or Pro), and sessions running through a cloud provider like Bedrock or Vertex can’t publish. Organizations on CMEK, HIPAA, or Zero Data Retention are excluded for now too.

TL;DR

Claude Code can now publish a coding session as a private, live, versioned web page — built from your repo, tools, and conversation — that updates in place and replaces the hand-written status update. It’s deliberately sealed: a capture of work, not an app. Use it to share; deploy for anything that has to persist or serve.