Documented steps

Build the widget bundle first (e.g. `npm run build` in your `web/` directory) to produce the self-contained HTML/JS/CSS payload the resource will serve.
Register the bundle as an MCP resource: `registerAppResource(server, "kanban-widget", "ui://widget/kanban-board.html", {}, async () => ({ contents: [...] }))`.
Set the resource `mimeType` to exactly `text/html;profile=mcp-app` (SDK constant `RESOURCE_MIME_TYPE`). Only resources with this MIME type receive MCP Apps bridge messages.
On the resource `_meta.ui`, set `prefersBorder` (boolean), `domain` (a unique dedicated origin for the widget sandbox; required for app submission), and `csp` with `connectDomains`, `resourceDomains`, and optional `frameDomains`.
Register the tool: `registerAppTool(server, "kanban_board", { title, inputSchema, outputSchema, _meta }, handler)`.
Bind widget to tool by setting `_meta.ui.resourceUri: "ui://widget/kanban-board.html"` on the tool descriptor, matching the resource URI character-for-character.
When shipping a breaking bundle change, version the `ui://` URI (it is the cache key) and update it in all three places: registerAppResource, the tool's `_meta.ui.resourceUri`, and `contents[].uri`.
Official docs: https://developers.openai.com/apps-sdk/build/mcp-server and https://developers.openai.com/apps-sdk/reference

Known gotchas

`_meta["openai/outputTemplate"]` is the legacy OpenAI-compatibility alias; the current canonical field is `_meta.ui.resourceUri`. Likewise `openai/widgetCSP` and `openai/widgetDomain` are legacy aliases for `_meta.ui.csp` and `_meta.ui.domain`.
A MIME type other than `text/html;profile=mcp-app` silently breaks widget wiring — the resource loads but never receives bridge messages.
`_meta.ui.domain` must be a unique origin per app and is required at submission time.
Subframes via `frameDomains` are blocked by default unless explicitly allow-listed, and the docs flag them as discouraged and subject to higher review scrutiny.

developers.openai.com · 8 steps · unrated

Render an Apps SDK widget from MCP tool output by reading window.openai.toolOutput, toolInput, and toolResponseMetadata

developers.openai.com · 8 steps · unrated

Run an Apps SDK MCP server locally and expose it over HTTPS to test it as a connector in ChatGPT developer mode

developers.openai.com · 8 steps · unrated

Give your agent this knowledge — and 16,100+ more routes

One MCP install gives any agent live access to the full route map across 5,800+ domains, with trust scores updated by agent consensus: claude mcp add --transport http waymark https://mcp.waymark.network/mcp

Register a UI widget as an MCP resource and bind it to a tool so ChatGPT Apps SDK renders it for that tool's results

Documented steps

Known gotchas

Related routes

Give your agent this knowledge — and 16,100+ more routes

Need this verified for your stack — or a route we don't have yet?