Can Claude Artifacts call AI from inside the app?

Yes. Since June 25, 2025, artifacts can call Claude directly through an in-pane API (documented as window.claude.complete). There's no separate API key and no per-call charge; usage counts against the user's normal plan limits, and the caller must be signed in.

How much data can a Claude Artifact store?

Up to 20 MB of persistent key-value storage per artifact on paid plans, shipped October 21, 2025. State is saved with the artifact and persists across sessions, but it is not portable to an external database.

What's the difference between publishing an Artifact and a Live Artifact?

Publishing creates a public shareable URL of a static-state app. A Live Artifact (launched April 20, 2026 in Cowork) connects to data through MCP connectors and auto-refreshes each time you reopen it, which is what makes dashboards stay current.

Can you share Claude Artifacts privately inside a company?

Yes, as of the June 18, 2026 Claude Code Artifacts beta for Team and Enterprise. It adds private organizational links scoped to your company, closing the gap where the public Publish button didn't work on Team and Enterprise plans.

What runtime do Claude Artifacts use?

They render in a sandboxed iframe on claudeusercontent.com with React 18, TypeScript, Tailwind, shadcn/ui, Lucide, Recharts, and Radix available by default. As of June 15, 2026, Team/Enterprise legacy artifacts migrated to a VM-based sandboxed container.

Claude Artifacts Quietly Became an App Platform

Somewhere between October 2025 and June 2026, Claude Artifacts stopped being a preview pane and became a place people ship software. Anthropic never ran a keynote calling it an app platform.

The capabilities just stacked up, one quiet release note at a time, until a chat side-panel could store data, call AI on its own, talk to Slack and Gmail, and hand you a private link to send a coworker.

By Anthropic's own count, 500 million artifacts had been created by June 25, 2025. That was before the features that actually made them apps had even shipped.

If you build with Claude and still think of Artifacts as "the thing that shows you a chart," this piece is the catch-up.

TL;DR

Claude Artifacts are now a no-backend app platform: a single conversation produces a working app with persistent storage, AI running inside it, and live connections to your tools, hosted and executed by Claude with no deployment step. Power users use it to ship micro-apps and dashboards in minutes, then graduate the few that matter to real infrastructure. The version-specific bits move fast, so anchor your workflow to the durable pattern, not this month's model.

Key takeaways

Three releases turned the pane into a platform: in-artifact AI calls (June 25, 2025), 20 MB persistent storage plus MCP (October 21, 2025), and private organizational links via Claude Code Artifacts (June 18, 2026).
The runtime is generous: React 18, TypeScript, Tailwind, shadcn/ui, Lucide, Recharts, and Radix are available without a build step.
Live Artifacts (April 20, 2026) auto-refresh from MCP connectors, which is why dashboards are the dominant 2026 build category.
The hard part is iteration, not generation. One-shot prompts give up; multi-prompt builders ship.
Know when to graduate. An artifact is not portable. Pull the winners out into real apps.

What makes Claude Artifacts an app platform now?

A Claude Artifact in 2026 is a self-contained interactive app, generated from a conversation, that can save state, call Claude's API from inside itself, connect to external services through MCP, and be shared by link, all without you provisioning a server or writing a deploy script.

That definition only became true through a specific sequence. Here's the spine.

Date	What shipped	Why it mattered
Jun 20, 2024	Artifacts launch with Claude 3.5 Sonnet	A preview pane for AI-generated content
Jun 25, 2025	In-artifact AI via `window.claude.complete()`	Apps could now think, not just render
Oct 21, 2025	20 MB persistent storage + MCP support	State across sessions, plus real data sources
Apr 20, 2026	Live Artifacts in Cowork	Auto-refreshing dashboards via MCP connectors
Jun 18, 2026	Claude Code Artifacts (Team/Enterprise beta)	Private org links, the missing enterprise piece

The June 25, 2025 release is the hinge. Anthropic's blog described it as artifacts that "run code, save state, and call external tools." Simon Willison reverse-engineered the in-pane API the same day, confirming what the community nicknamed "Claudeception": an artifact can send a prompt to Claude and get a completion back.

The verbatim system-prompt text notes the API returns a string by default, so you instruct Claude to return JSON if you want structured output.

The October 21, 2025 update added the two things every real app needs: memory and connections. Persistent storage of up to 20 MB per artifact on paid plans, plus MCP connectors for Slack, Google Calendar, Gmail, Notion, and Linear. State saves with the artifact and survives across sessions.

What do power users actually build with Claude Artifacts?

Dashboards, mostly. They're the dominant 2026 category because Live Artifacts made them stay current on their own.

Kunal Ganglani built a dark drone fleet management dashboard with live telemetry, geospatial mapping, battery status, and WCAG AA contrast in 14 prompts over 90 minutes, wired to a mock telemetry source through an MCP connector. That ratio, double-digit prompts and roughly an hour-and-a-half, is a realistic benchmark for a polished build, not a one-line miracle.

Nuno Roberto built a working Kanban board during a meeting: status columns, drag-and-drop cards, a live progress bar. His quote captures the moment well, a colleague asking "How did you do that while I was still talking?"

Then there's the long tail. Simon Willison shipped 14 tools in a single week as artifacts: a URL-to-Markdown converter, a SQLite-WASM demo, a Pyodide REPL, an LLM pricing calculator, a QR decoder. His documented trick is a "no react" prefix that gets Claude to emit portable single-file HTML/JS he drops straight into GitHub Pages.

The "starter menu" that recurs across 2026 community galleries: sales quote calculators, KPI dashboards, content calendars, habit trackers, decision matrices, quiz lead-magnets, resume builders. Most are ten-minute builds.

The runtime is more capable than people expect

A big reason complex artifacts "just work" is the default environment. Willison confirmed the stack by running the community claude-artifact-runner: React 18, TypeScript, Vite/Parcel, Tailwind, shadcn/ui (40+ components), Lucide icons, Recharts, and Radix primitives, all available without bundling.

Anthropic packaged this as the web-artifacts-builder skill, which compiles that stack into a single self-contained HTML file and has logged 55,825 downloads. That skill is why a drag-and-drop Kanban appears in a side panel instead of erroring out on a missing dependency.

Everything renders inside a sandboxed iframe on claudeusercontent.com with strict CSP headers, per the best technical writeup of the model. There's no filesystem access, and fetch() is CSP-restricted unless the target is CORS-allowed, which is exactly why the in-pane Claude call and MCP connectors exist as the sanctioned escape hatches.

Live Artifacts and private links: the 2026 enables

Live Artifacts launched April 20, 2026 in Cowork. The distinction from a published artifact is simple. A Live Artifact refreshes its data on a schedule or on each open through MCP connectors, with version history in a sidebar tab on Pro+ plans. That's the difference between a screenshot of a dashboard and a dashboard.

The enterprise gap was distribution. The public Publish button, as send.co documented, doesn't work on Team/Enterprise plans, gives no view counts, and makes unpublishing permanent. The June 18, 2026 Claude Code Artifacts beta closes it with private organizational links scoped to the company.

Anthropic's three highlighted patterns are pointed and high-leverage: annotated PR walkthroughs, session-data dashboards, and investigation timelines for agent runs that went off the rails.

The real limit: generation outruns iteration

The most useful framing of the moment comes from Siqi Li's May 22, 2026 essay, "Claude Artifacts and the Second Shockwave of AI Design." Her central tension, paraphrased into a line worth keeping: generation capability is advancing much faster than iteration efficiency.

Claude can scaffold a polished UI in one prompt. Getting it to shipped quality still takes rounds. The people who give up are the ones trying to one-shot a finished product. The people who ship treat it as a 14-prompt conversation, the way Ganglani did.

Li also names the UI honestly. The artifact pane "feels like an embedded widget, not native." It's a sibling to chat, not a primary surface. That's the practical reason power users either learn to ship via links or graduate to a real app.

When should you graduate an Artifact to production?

When it stops being a toy and starts being load-bearing. Lily Luo's productionization framework is the clearest guide of the year, and it gives you three exits:

Embed in an existing workflow (a CRM button, a SharePoint folder, a Slack app) when the artifact is useful but lives next to where people already work.
Promote to a dedicated micro-app (Next.js, Vercel, Postgres) when it needs real auth, a real database, or real uptime.
Build an agent on top when the artifact's surface is the front-end for something more autonomous.

The bridge problem she names: an artifact in your chat is not portable. The leverage comes from pulling the AI out of the chat and into the actual workflow.

Watch the sandbox, not just the build

Two security facts belong in any practitioner's mental model. First, because the iframe can call window.claude.complete, a malicious artifact can act as a confused deputy, inheriting your signed-in Claude session and permissions.

Anthropic's "Trustworthy Agents in Practice" framework (April 2026) names this class of risk. Second, public links carry public-link risk: researchers tracked a Mac malware campaign that abused Google Ads pointing at benign-looking public Claude Artifact URLs.

Treat a public artifact link the way you'd treat any untrusted page. The workaround is boring and effective: prefer private organizational links for anything internal, and don't load artifacts from sources you don't trust.

Note the runtime shift too. As of June 15, 2026, Anthropic retired legacy Team/Enterprise artifacts in favor of VM-based artifacts in a sandboxed container, closer to a V8 isolate than a browser iframe. Admins had to enable code execution. If your org's artifacts broke that week, that's why.

What this means for you

Start building dashboards and internal tools in Artifacts before you reach for a framework. The fastest path to a working spend tracker, KPI board, or quiz is a multi-prompt conversation, not a repo.

Plan for iteration. Budget 10-plus prompts for anything polished, and use the in-pane AI call for the "smart" parts (scoring feedback, drafting replies, summarizing) instead of bolting on a separate service.

Connect data through MCP and Live Artifacts when the thing needs to stay current, and reach for private organizational links the moment a build leaves your own screen.

Graduate deliberately. Most artifacts should stay artifacts. The few that become load-bearing deserve a real Next.js app and a real database.

One durable caveat on versions: Anthropic ships models fast (Opus 4.8 landed May 28, 2026; Sonnet 4.6 on February 17, 2026), so anchor your workflow to the platform capabilities here, not to whichever model is default this week. The capabilities are sticky. The version numbers rot in weeks.

What to watch next: whether the VM-based runtime opens up outbound network access beyond MCP, and whether private links pick up the view counts and access controls that would make Artifacts a genuine internal app store.

Claude Artifacts Quietly Became a No-Backend App Platform