How to give your own AI agent Claude-style Artifacts

An agent that can write HTML but can only paste it into a chat is missing the most useful part of an artifact: a page someone can open. To give your own AI agent Claude-style Artifacts, add one publishing tool to its workflow. OpenArtifacts is built for this: an agent can sign up in a few seconds, publish a page, and return a live URL. A human can sign up in seconds too, or claim the account the agent started.

Claude Artifacts already makes designs, documents, dashboards, and interactive tools inside Claude. Claude also has artifact sharing. If you are looking for a Claude Artifacts alternative for an agent you built, OpenArtifacts gives that agent a way to publish its output to the web from whichever model and interface you choose.

The four pieces your agent needs

1. A page to publish. Have the agent write a complete index.html, with relative links if it also writes CSS or JavaScript files.
2. A publish operation. Give the agent the OpenArtifacts skill, its remote MCP tools, or an HTTP tool that calls the REST API.
3. A stable URL. Return the url from the publish response to the human. Reusing the same slug for an edit creates a new version at that URL.
4. A human owner. Return the claim_url as well. The human opens it and completes the claim within 30 minutes to keep the account and its pages.

That is the artifact loop: generate, publish, share, revise. Your agent does not need to build an account system or a hosting service first.

Start with two HTTP calls

The agent signs up without a form or email address:

curl -sX POST https://api.openartifacts.sh/v1/signup

The response contains a token and a claim_url. Save the token privately; it is shown once. For a first page, use it in the next call:

curl -sX POST https://api.openartifacts.sh/v1/publish \
  -H "Authorization: Bearer $OA_TOKEN" \
  -F slug=weekly-dashboard \
  --form-string html='<!doctype html><html lang="en"><meta charset="utf-8"><title>Weekly dashboard</title><h1>Weekly dashboard</h1></html>'

Read the returned url rather than constructing one. Send it to the human with the claim_url from signup. The agent can publish immediately on a provisional account; the human must complete the claim within 30 minutes or the account and its pages expire. A person who wants to start first can create a human account with a username and password, then create an agent token in the dashboard. Both paths take only a few steps and cost nothing to start. Claiming unlocks 100 free publishes per month.

Put the capability in your agent

For an MCP-capable agent, connect https://mcp.openartifacts.sh/mcp with the token as a bearer authorization header. The server exposes publish_artifact for text files, plus tools to list, inspect, and roll back artifacts. Our MCP setup guide includes a Claude Code example; other MCP clients use their own connection syntax.

For an agent with an HTTP or shell tool, give it a short instruction it can act on:

When the user asks for a shareable page, write the HTML and any local assets, publish them with OpenArtifacts, and return the exact url. Reuse the slug when revising the page. If you signed up for a provisional account, return claim_url in the same message and explain its 30-minute deadline. Keep the bearer token private.

If the agent produces a folder rather than one HTML string, upload index.html, CSS, JavaScript, and images as multipart files. The multi-file guide shows the request and explains relative paths. Publishing the same slug creates a new version, so the link you already shared keeps working; you can activate an earlier version if needed.

Where this fits

OpenArtifacts gives an agent you control a ready-made publishing layer: send it HTML, CSS, JavaScript, and assets over an API or MCP, and get a public URL back. The same flow works whether your agent uses Claude, another model, a terminal, or your own application. If you want to give that agent the equivalent of a shareable artifact, start with the OpenArtifacts skill and let it publish one page.

Building a similar workflow around OpenAI's hosting feature? See How to give your own agent ChatGPT Sites-style publishing.