Claude can now build apps that connect to your real tools: calendar, tasks, analytics. Most people have no idea and still use it for chat.
Inside: how artifacts actually work, the new connection setup, and 4 tools worth building this week.
Learn this once and you stop paying for half the software you use.
The 5 levels of artifacts
Artifacts stopped being one feature a while ago. It's a ladder, and each level unlocks a different class of tools:
Level 1, static output: documents, diagrams, SVG, single-page designs. Available on every plan including FreeLevel 2, interactive apps: React components and HTML pages with working buttons, forms, and state. Still just a smarter previewLevel 3, persistent storage: artifacts remember data between sessions, up to 20MB per artifact. Now your habit tracker doesn't reset on refreshLevel 4, Claude inside: an artifact can call the Claude API directly. People call it "Claude in Claude": the app you built has AI running inside itLevel 5, MCP connectors (new): artifacts read and write your real services. A dashboard that pulls live BigQuery data, a form that creates Asana tasksLevels 3-5 need Pro, Max, Team, or Enterprise, on web and desktop.
Building your first real artifact
Skip "make me a website". The prompt pattern that produces working tools:
Build an artifact: [what it is] for [who uses it].
Core features: [3-5 specific functions]
Data: [what it stores or fetches]
Design: [style reference or "clean, minimal"]
Use persistent storage so data survives sessions.
Build an artifact: a content calendar for a solo creator.
Core features: add post ideas with status (draft/scheduled/posted),
filter by platform, weekly view
Data: store all entries persistently
Design: clean, dark mode
One prompt, working tool. Then iterate: "add a stats bar showing posts per week" instead of re-describing everything.
The new part: connecting artifacts to real services
This is what shipped yesterday. The flow, verified from the docs:
Build or open an artifact that needs external data ("dashboard showing my calendar and open GitHub issues")On first interaction, the artifact shows a connector approval prompt: "This artifact uses connectors" with a Connect button per serviceApprove access. Your choice persists for that artifact going forwardThe artifact now fetches and writes data on demand, live, per viewerThree rules that decide whether your build works:
Each viewer authenticates separately. Share an artifact with your team, and every person connects their own accounts. Your GitHub token never leaks to a teammate, and their dashboard shows their dataPublicly-shared artifacts can't use connectors. Live-data apps stay private or org-internal. Plan your builds accordinglyAdmins hold the switch. On Team/Enterprise, org admins can disable artifact MCP access entirely, so check before promising your team a tool4 builds actually worth making
Personal ops dashboard. Calendar + task list + inbox count in one tab. The classic "what should I focus on today" screen, now with live data instead of pasted screenshotsClient-facing report page. Pulls metrics from your analytics connector each time it opens. Build once, stop exporting PDFs every MondayTeam intake form. A form that writes directly into Asana or Linear through the connector, with your validation rules on top. No Zapier taxLive Artifacts move: persistent dashboards live in their own tab, refresh with current data each time you open them, and keep version history. Treat them as your personal internal tools, not chat outputsPublishing and remixing
The distribution layer most people never touch:
Publish any artifact via link. Viewers don't need a Claude account for static and interactive levelsRemixing: anyone with a Claude account can fork a published artifact into their own copy and modify it. Your tool becomes a templateCommunity catalog: a browsable library of published artifacts. Before building from scratch, check if someone shipped 80% of your idea alreadyCommon mistakes
Building connector apps for public sharing. Connectors don't work on public artifacts. If the goal is a public tool, design around static data or the Claude API level insteadRe-describing the whole app on every edit. Artifacts iterate. Ask for one change at a time and the quality stays consistentIgnoring the 20MB storage ceiling. Persistent storage is for state and entries, not media libraries. Keep images externalExpecting shared artifacts to share your data. Per-user auth means every viewer sees their own services. That's a feature for teams and a surprise for the unpreparedSkipping the approval prompt details. The connect dialog names exactly which services the artifact touches. Read it, especially on remixed artifacts someone else builtThe 15-minute setup
Pick one repetitive check you do daily: calendar, metrics, tasks (1 min)Prompt the artifact with the pattern above, persistent storage on (3 min)Add the connector, approve access on first run (2 min)Iterate twice: one layout fix, one feature add (7 min)Pin the tab. That's your first internal tool (2 min)The chat window was never the product. The things you build in it are.