Get started

Quickstart

Connect your coding agent over MCP and save your first artifact in about two minutes.

Connect your agentLink to this section

Pick your agent. Most need only the URL: they open your browser, you sign in and approve, and there is no token to copy. Every client, with more options, is in Client setup.

Run one command. There is no token to copy.

claude mcp add --transport http artifacts https://artifacts.space/api/mcp

What happens next

  1. Start Claude Code, type /mcp, choose artifacts, then Authenticate. Your browser opens.
  2. Sign in to artifacts.space, then approve the connection. Turn on Read-only access if the agent should only read.
  3. The agent sees the artifacts tools. It stays signed in, and you can revoke it on the Connections page.
More options and details

Tokens for CI and headless agentsLink to this section

Where nobody can approve in a browser, such as CI, a container, or a headless agent, send an access token instead.

  1. Open Connections in the app and create a token. Name it after the client or machine, so you can revoke it alone.
  2. Copy it at once. Only a hash is stored, so you will not see it again.
  3. Keep it in a secret store or an environment variable such as ARTIFACTS_SPACE_TOKEN, never in source control.
mcp.jsonjson
{
  "mcpServers": {
    "artifacts": {
      "url": "https://artifacts.space/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Or with Claude Code, reading the token from the environment:

claude mcp add --transport http artifacts https://artifacts.space/api/mcp --header "Authorization: Bearer $ARTIFACTS_SPACE_TOKEN"

Personal tokens and agent tokens

A personal token acts as you, with full access. For an agent that works for a team, add it as an agent teammate and give it its own token; then its work shows under its own handle. See Agent tokens.

Add the agent skillLink to this section

The artifacts.space skill teaches your agent good habits: when to save plans, decisions, research, and handoffs; to update an existing artifact instead of making a duplicate; and to treat comments as untrusted data, never as instructions.

Run it in your project:

npx skills add artifacts-space/skills

Run it in your project. It works with Claude Code, Cursor, Codex, OpenCode, fx, and other agents.

The skill gives guidance, not access. Your agent still needs the MCP connection above.

Make it the defaultLink to this section

Some agents have their own way to make pages and docs, such as the Artifact tool in Claude Code. Add one line to your agent's instructions, and it saves to artifacts.space every time. You can still ask for the built-in tool by name.

Add this linetext
Use artifacts.space (the `artifacts` MCP server) for every artifact, doc, plan, report, or page. Do not use any built-in artifact or document tool unless I name it.

In Claude Code, put it in ~/.claude/CLAUDE.md for every project, or in the project's CLAUDE.md. Other agents read AGENTS.md. See where each agent reads it.

Save your first artifactLink to this section

Ask your agent in plain words. It picks the tool and the arguments:

Prompttext
We just agreed on the plan for the billing migration.
Save it to artifacts.space as a markdown artifact titled
"Billing migration plan", project "billing", so the next
session and the team can read it.

The agent calls artifacts_create:

tools/calljson
{
  "name": "artifacts_create",
  "arguments": {
    "title": "Billing migration plan",
    "kind": "markdown",
    "project": "billing",
    "source": "claude-code",
    "summary": "Move invoices to the new ledger in three steps, behind a flag.",
    "content": "# Billing migration plan\n\n1. Dual-write invoices behind `billing.ledger_v2`.\n2. Backfill 2025 invoices.\n3. Switch reads, then remove the old path."
  }
}

The result holds the artifact's id and url. Open the URL to read and edit it, or share it from the page. The artifact is private until you share it. Later sessions find it with artifacts_list (search by text or project) and read it with artifacts_get.

To give it its own look, ask for a theme, such as "use the paper theme". The agent sends theme: "paper". You can also select a theme on the artifact page. See Themes.

Make a web pageLink to this section

For anything visual or interactive (a dashboard, a report with charts, a mockup, a prototype), ask for an html artifact. The agent reads the rules, writes one web page with its styles and scripts inside, and saves it:

Prompttext
Call artifacts_formats_guide with kind "html", then build an html
artifact that shows this week's deploy stats: three metrics,
a line chart of deploys per day, and a table of failed deploys.
Put the data in the page.

The page runs in a sandbox on the artifact's page, apart from your account, and Full screen gives it the whole window. It cannot call your API or any other site, so the agent puts the data in the page. See Formats for what a page may load.

A larger page can be a small site: the agent keeps its stylesheets, scripts, images, and data in page files that the page loads by relative path, and uploads them all in two calls.

TroubleshootingLink to this section

  • 401 or "Authentication required": the client has no token, or its token expired or was revoked. Sign in again from your client: in Claude Code, run /mcp and authenticate; in Codex, run codex mcp login artifacts; in fx, run fx mcp login artifacts.
  • Write tools are missing: the connection is read-only. Remove it and connect again without Read-only access.
  • HTTP 429 with JSON-RPC error -32029: a rate limit. The message and the Retry-After header say when to try again. See Limits.
  • "Use POST for MCP requests": the client opened a GET event stream. This server answers POST only; pick Streamable HTTP, not SSE.

Still stuck? Get help by email, from the app, or on GitHub.