Claude Code Artifacts: Instantly Deliver and Share Result Pages

Claude Code has incorporated the previously web‑only artifacts feature, turning conversation outputs into private, continuously accessible web pages with unique URLs that do not interfere with the ongoing session.

Artifacts act as deliverables: Claude generates an HTML or Markdown file based on a plain‑language description (e.g., "create a PR review page with annotations"), publishes it, and asks for confirmation on the first release; subsequent updates are automatic.

When natural language does not trigger artifact creation, the /artifacts-design command can explicitly set the current session’s deliverable.

Typical uses highlighted include a PR review page with diff rendering and side‑panel annotations, a multi‑layout comparison grid with trade‑off notes, and an interactive parameter‑tuning page where sliders adjust animation parameters in real time.

The "bring the result back" pattern lets users drag cards on an issue triage board, copy the final ordering as a prompt, and feed it back into the conversation, creating a closed loop between UI actions and the chat.

Long‑running task tracking can be built as a checklist artifact that updates progress as Claude marks items, with collaborators viewing the live status via the shared link.

Artifacts are not full applications: each page is a single self‑contained HTML document running under a strict Content Security Policy (CSP). External scripts, stylesheets, and fetch calls are blocked; CSS and JavaScript must be inlined, images are embedded as data URIs, the rendered size is capped at 16 MB, and forms or multi‑page routing are unavailable.

Claude Design, announced earlier by Anthropic, is a separate design‑focused product that creates prototypes and handoff bundles for Claude Code; while both produce pages, Design offers a canvas and export capabilities that artifacts lack.

The latest update adds per‑viewer permission handling for MCP connectors: when an artifact pulls data (e.g., a GitHub dashboard), the request runs under the viewer’s account, not the publisher’s, so different users may see different data based on their own repository permissions. The first access prompts each viewer to grant permission, and no credentials are exposed in the page.

Because actions run under the viewer’s identity, artifacts can embed buttons that trigger side‑effects such as sending messages or updating issues, effectively turning a shared page into a lightweight front‑end with individualized permissions.

To enable MCP connector calls in a newly created artifact, the /artifact-capabilities command is used to declare the required capabilities.

Interested readers can try the artifacts feature themselves.