Local Artifacts

English | 简体中文

A local, open-source look‑alike of Claude Code's Artifacts feature — for personal accounts that can't use the official one.

License: MIT

Turn Claude Code's work into a live, interactive web page at a local URL that updates in place as your session continues.

This is an independent local re-implementation of an Artifacts-style workflow. It does not access, unlock, modify, or proxy Anthropic's official Artifacts service — it only renders content your own local Claude Code session produces, on your own machine.

You: "把这个分析做成 artifact 发布"   →   Claude generates a page
     →  publishes to http://localhost:7891  →  browser opens
     →  say "更新一下" → page reloads in place (SSE live refresh)

⚠️ Unofficial — read this first

This project is NOT Anthropic's official Artifacts feature and is not affiliated with Anthropic. It only imitates the user-facing experience. The internals are fundamentally different:

If you have a Team/Enterprise plan, use the official feature instead — it's better and properly sandboxed.

What this is good for: personal accounts, API-key users, or anyone who just wants "make my Claude Code session output a shareable local page."

Features

• 🎯 One command — say "make an artifact" in any Claude Code session
• 🔄 Live refresh — same URL updates in place via Server-Sent Events
• 📝 HTML + Markdown — Markdown rendered by bundled marked.js (no CDN; works fully offline / on intranets)
• 🔒 XSS-hardened — </script> escaping + HTML-escaped titles
• 💾 Persistent — reopens your last artifact after a restart
• 📦 Self-contained — one Python file + one JS file, only dependency is mcp

Requirements

• macOS / Linux / Windows
• Python 3.10+
• Claude Code
• Python package: mcp

Install

git clone https://github.com/xiagaohui/local-artifacts-for-claude-code.git
cd local-artifacts-for-claude-code
./install.sh

The script will:

1. Create a virtualenv and install mcp
2. Register the local-artifacts MCP server with Claude Code
3. Install the /artifact skill

Then restart Claude Code. Done.

# 1. install the mcp dependency into a venv
python3 -m venv .venv
.venv/bin/pip install mcp

# 2. register the MCP server (adjust the absolute path)
claude mcp add local-artifacts \
  "$(pwd)/.venv/bin/python" "$(pwd)/server.py"

# 3. (optional) install the /artifact reopen skill
mkdir -p ~/.claude/skills/artifact
cp skill/SKILL.md ~/.claude/skills/artifact/SKILL.md

# 4. restart Claude Code

Usage

In any Claude Code session:

把这个结果做成 artifact 发布            # publish as an artifact
帮我做一个 PR 走查 artifact，展示 diff   # PR walk-through
做一个数据看板 artifact                  # dashboard
更新一下 artifact，加上 xxx              # update in place

Reopen the latest artifact anytime:

/artifact

Configuration

How it works

Claude Code ──(MCP stdio)──> server.py ──┬── publish_artifact tool
                                          └── ThreadingHTTPServer :7891
                                                ├── GET /          → rendered page
                                                ├── GET /events    → SSE live-reload
                                                └── POST /publish   → update content
Browser ──> http://localhost:7891 ──(EventSource)──> auto-reload on publish

Security

Please read before installing:

• It renders arbitrary HTML/JS. An artifact is a page built from content your Claude Code session produces, served at http://localhost:7891. Opening it executes that HTML/JavaScript in your browser — treat it like running untrusted front-end code. Only publish content you trust.
• Localhost only. The server binds to 127.0.0.1, so it is not reachable from your LAN/Wi-Fi. Don't expose the port (e.g. via a tunnel) to untrusted networks.
• file_path is restricted. The publish_artifact tool only reads document files (.html/.htm/.md/.markdown/.txt); it refuses other paths to avoid leaking files like ~/.ssh/id_rsa or .env if a prompt tries to make it.
• Local state is private. Published content is cached under ~/.claude/artifacts/ with 0600 permissions (dir 0700).
• No sandbox. Unlike the official feature (CSP-isolated), this has no content sandbox. That's the price of running it yourself on a personal account.

Use it for your own work on your own machine. Don't point it at content from untrusted sources.

Limitations vs. official

Intentionally omitted (not needed for personal single-machine use): org sharing, version history & picker, author gallery, audit logs, CSP sandboxing, retention policies, Compliance API. See the comparison table above.

Credits

• Markdown rendering by marked (MIT) — bundled as marked.min.js.
• Inspired by Claude Code Artifacts.

License

MIT