This skill should be used when the user wants to build an "MCP app", add "interactive UI" or "widgets" to an MCP server, "render components in chat", build "MCP UI resources", make a tool that shows a "form", "picker", "dashboard" or "confirmation dialog" inline in the conversation, or mentions "apps SDK" in the context of MCP. Use AFTER the build-mcp-server skill has settled the deployment model, or when the user already knows they want UI widgets.
This skill should be used when the user wants to build an "MCP app", add "interactive UI" or "widgets" to an MCP server, "render components in chat", build "MCP UI resources", make a tool that shows a "form", "picker", "dashboard" or "confirmation dialog" inline in the conversation, or mentions "apps SDK" in the context of MCP. Use AFTER the build-mcp-server skill has settled the deployment model, or when the user already knows they want UI widgets.
version
0.1.0
Build an MCP App (Interactive UI Widgets)
An MCP app is a standard MCP server that also serves UI resources — interactive components rendered inline in the chat surface. Build once, runs in Claude and ChatGPT and any other host that implements the apps surface.
The UI layer is additive. Under the hood it's still tools, resources, and the same wire protocol. If you haven't built a plain MCP server before, the build-mcp-server skill covers the base layer. This skill adds widgets on top.
Testing in Claude: Add the server as a custom connector in claude.ai (via a Cloudflare tunnel for local dev) — this exercises the real iframe sandbox and hostContext. See https://claude.com/docs/connectors/building/testing.
Claude host specifics
_meta.ui.* key
Where
Effect
resourceUri
tool
Which ui:// resource the host renders for this tool's results.
visibility: ["app"]
tool
Hide a widget-only helper tool (e.g. geometry/image fetcher called via callServerTool) from Claude's tool list.
Declare external origins; default is block-all. frameDomains is currently restricted in Claude.
hostContext.safeAreaInsets: {top, right, bottom, left} (px) — honor these for notches and the composer overlay.
Directory submission requires OAuth or authless (none) — static bearer is private-deploy only and blocks listing — plus tool annotations and 3–5 PNG screenshots; see references/directory-checklist.md.
When a widget beats plain text
Don't add UI for its own sake — most tools are fine returning text or JSON. Add a widget when one of these is true:
Signal
Widget type
Tool needs structured input Claude can't reliably infer
Form
User must pick from a list Claude can't rank (files, contacts, records)
Picker / table
Destructive or billable action needs explicit confirmation
Confirm dialog
Output is spatial or visual (charts, maps, diffs, previews)
Display widget
Long-running job the user wants to watch
Progress / live status
If none apply, skip the widget. Text is faster to build and faster for the user.
Widgets vs Elicitation — route correctly
Before building a widget, check if elicitation covers it. Elicitation is spec-native, zero UI code, works in any compliant host.
Need
Elicitation
Widget
Confirm yes/no
✅
overkill
Pick from short enum
✅
overkill
Fill a flat form (name, email, date)
✅
overkill
Pick from a large/searchable list
❌ (no scroll/search)
✅
Visual preview before choosing
❌
✅
Chart / map / diff view
❌
✅
Live-updating progress
❌
✅
If elicitation covers it, use it. See ../build-mcp-server/references/elicitation.md.
Architecture: two deployment shapes
Remote MCP app (most common)
Hosted streamable-HTTP server. Widget templates are served as resources; tool results reference them. The host fetches the resource, renders it in an iframe sandbox, and brokers messages between the widget and Claude.
Same widget mechanism, but the server runs locally inside an MCPB bundle. Use this when the widget needs to drive a local application — e.g., a file picker that browses the actual local disk, a dialog that controls a desktop app.
For MCPB packaging mechanics, defer to the build-mcpb skill. Everything below applies to both shapes.
How widgets attach to tools
A widget-enabled tool has two separate registrations:
The tool declares a UI resource via _meta.ui.resourceUri. Its handler returns plain text/JSON — NOT the HTML.
The resource is registered separately and serves the HTML.
When Claude calls the tool, the host sees _meta.ui.resourceUri, fetches that resource, renders it in an iframe, and pipes the tool's return value into the iframe via the ontoolresult event.
The URI scheme ui:// is convention. The mime type MUST be RESOURCE_MIME_TYPE ("text/html;profile=mcp-app") — this is how the host knows to render it as an interactive iframe, not just display the source.
Widget runtime — the App class
Inside the iframe, your script talks to the host via the App class from @modelcontextprotocol/ext-apps. This is a persistent bidirectional connection — the widget stays alive as long as the conversation is active, receiving new tool results and sending user actions.
<scripttype="module">/* ext-apps bundle inlined at build time → globalThis.ExtApps *//*__EXT_APPS_BUNDLE__*/const { App } = globalThis.ExtApps;
const app = newApp({ name: "ContactPicker", version: "1.0.0" }, {});
// Set handlers BEFORE connecting
app.ontoolresult = ({ content }) => {
const contacts = JSON.parse(content[0].text);
render(contacts);
};
await app.connect();
// Later, when the user clicks something:functiononPick(contact) {
app.sendMessage({
role: "user",
content: [{ type: "text", text: `Selected contact: ${contact.id}` }],
});
}
</script>
The /*__EXT_APPS_BUNDLE__*/ placeholder gets replaced by the server at startup with the contents of @modelcontextprotocol/ext-apps/app-with-deps — see references/iframe-sandbox.md for why this is necessary and the rewrite snippet. Do notimport { App } from "https://esm.sh/..."; the iframe's CSP blocks the transitive dependency fetches and the widget renders blank.
Method
Direction
Use for
app.ontoolresult = fn
Host → widget
Receive the tool's return value
app.ontoolinput = fn
Host → widget
Receive the tool's input args (what Claude passed)
app.sendMessage({...})
Widget → host
Inject a message into the conversation
app.updateModelContext({...})
Widget → host
Update context silently (no visible message)
app.callServerTool({name, arguments})
Widget → server
Call another tool on your server
app.openLink({url})
Widget → host
Open a URL in a new tab (sandbox blocks window.open)
sendMessage is the typical "user picked something, tell Claude" path. updateModelContext is for state that Claude should know about but shouldn't clutter the chat. openLink is required for any outbound navigation — window.open and <a target="_blank"> are blocked by the sandbox attribute.
What widgets cannot do:
Access the host page's DOM, cookies, or storage
Make network calls to arbitrary origins (CSP-restricted — route through callServerTool)
Open popups or navigate directly — use app.openLink({url})
Load remote images reliably — inline as data: URLs server-side
Keep widgets small and single-purpose. A picker picks. A chart displays. Don't build a whole sub-app inside the iframe — split it into multiple tools with focused widgets.
For local-only widget apps (driving a desktop app, reading local files), swap the transport to StdioServerTransport and package via the build-mcpb skill.
See references/widget-templates.md for more widget shapes.
Design notes that save you a rewrite
One widget per tool. Resist the urge to build one mega-widget that does everything. One tool → one focused widget → one clear result shape. Claude reasons about these far better.
Tool description must mention the widget. Claude only sees the tool description when deciding what to call. "Opens an interactive picker" in the description is what makes Claude reach for it instead of guessing an ID.
Widgets are optional at runtime. Hosts that don't support the apps surface simply ignore _meta.ui and render the tool's text content normally. Since your tool handler already returns meaningful text/JSON (the widget's data), degradation is automatic — Claude sees the data directly instead of via the widget.
Don't block on widget results for read-only tools. A widget that just displays data (chart, preview) shouldn't require a user action to complete. Return the display widget and a text summary in the same result so Claude can continue reasoning without waiting.
Layout-fork by item count, not by tool count. If one use case is "show one result in detail" and another is "show many results side-by-side", don't make two tools — make one tool that accepts items[], and let the widget pick a layout: items.length === 1 → detail view, > 1 → carousel. Keeps the server schema simple and lets Claude decide count naturally.
Put Claude's reasoning in the payload. A short note field on each item (why Claude picked it) rendered as a callout on the card gives users the reasoning inline with the choice. Mention this field in the tool description so Claude populates it.
Normalize image shapes server-side. If your data source returns images with wildly varying aspect ratios, rewrite to a predictable variant (e.g. square-bounded) before fetching for the data-URL inline. Then give the widget's image container a fixed aspect-ratio + object-fit: contain so everything sits centered.
Follow host theme.app.getHostContext()?.theme (after connect()) plus app.onhostcontextchanged for live updates. Toggle a .dark class on <html>, keep colors in CSS custom props with a :root.dark {} override block, set color-scheme. Disable mix-blend-mode: multiply in dark — it makes images vanish.
Testing
Claude Desktop — current builds still require the command/args config shape (no native "type": "http"). Wrap with mcp-remote and force http-only transport so the SSE probe doesn't swallow widget-capability negotiation:
Desktop caches UI resources aggressively. After editing widget HTML, fully quit (⌘Q / Alt+F4, not window-close) and relaunch to force a cold resource re-fetch.
Headless JSON-RPC loop — fast iteration without clicking through Desktop:
# test.jsonl — one JSON-RPC message per line
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"t","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/list"}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"your_tool","arguments":{...}}}
(cat test.jsonl; sleep 10) | npx mcp-remote http://localhost:3000/mcp --allow-http
The sleep keeps stdin open long enough to collect all responses. Parse the jsonl output with jq or a Python one-liner.
Widget dev loop — avoid the ⌘Q-relaunch cycle entirely by serving the inlined widget HTML at a plain GET route with a fake ExtApps shim that fires ontoolresult from a query param:
Open http://localhost:3000/widget-preview?payload={"rows":[...]} in a normal browser tab and iterate with ordinary devtools.
Host fallback — use a host without the apps surface (or MCP Inspector) and confirm the tool's text content degrades gracefully.
CSP debugging — open the iframe's own devtools console. CSP violations are the #1 reason widgets silently fail (blank rectangle, no error in the main console). See references/iframe-sandbox.md.
