Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What are MCP Apps?

MCP Apps is an official extension to the Model Context Protocol that lets an MCP server return an interactive HTML interface, such as a chart, form or viewer, which the chat host renders inline in a sandboxed iframe. The app talks to the host over JSON-RPC via postMessage and can call the server's tools itself.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

The host calls a tool, fetches the ui:// resource it points at, renders it in a sandboxed iframe, and relays the app's own tool calls back to the server.you, in the chatchat hostClaude, VS Code, GooseMCP Appsandboxed iframeMCP server1. tools/call2. result + ui:// HTML3. renders, pushes data4. clicks, types5. tools/call, postMessage6. forwards the call
  1. The server registers a tool whose _meta.ui.resourceUri points at a resource with the ui:// scheme, served with the type text/html;profile=mcp-app.

    _meta: { ui: { resourceUri: "ui://charts/bar" } }
  2. When the model calls the tool, the host fetches that resource, puts the HTML in a sandboxed iframe, and pushes the tool result into it.

  3. Inside the frame the app sends ui/initialize, then talks to the host over postMessage with JSON-RPC: tools/call, ui/open-link, ui/update-model-context.

  4. The host applies a Content Security Policy from the resource's _meta.ui.csp. Declare nothing and the default blocks every outside origin.

  5. A host without the extension ignores the UI and shows the tool's text content, so every UI tool still has to return meaningful text.

Source: MCP extensions: MCP Apps

MCP Apps versus plain tool results

A normal MCP tool returns content the host prints into the conversation: text, an image, a resource link. The model reads it and the person reads it, and that is the end. MCP Apps keeps the tool but adds a second half: a resource with a ui:// address that holds an HTML page. The tool's _meta.ui.resourceUri ties the two together. When a host that supports the extension sees that field, it fetches the page, renders it in a sandboxed iframe in place of the plain result, and hands the result data to the page.

Two design choices matter more than the rest. First, the page is a predeclared resource, not something the tool generates on the fly, so the host can fetch and inspect it before the tool ever runs, and cache it. Second, the app and the host speak the same JSON-RPC the rest of MCP uses, carried over postMessage instead of stdio or HTTP. Some methods are shared with core MCP (tools/call, resources/read), some are specific to apps (ui/initialize, ui/open-link, ui/update-model-context). Every message can be logged, and the host can refuse or ask for approval before forwarding a call.

Example: a chart you can click

A server that reports sales has a sales_by_region tool. Without the extension, the tool returns a table as text. With it, the tool's metadata points at ui://sales/map and the resource serves a single HTML file, scripts bundled in, with the type text/html;profile=mcp-app. You ask Claude for sales by region; the host calls the tool, gets both the numbers and the map, and renders the map in the chat. You click a region. The app calls sales_by_region again with that region as an argument, through the host, and redraws. The host saw both calls and could have asked you before the second one.

If you then switch to a client without the extension, the same tool runs and you get the table. The spec asks servers to keep that text fallback meaningful, and this is why.

When to use an MCP App

Use one when the answer is something a person wants to poke at rather than read: a dashboard with filters, a form with twenty interdependent fields, a PDF or 3D model viewer, a queue of items to approve one at a time. Do not use one for a result that is a sentence. The iframe costs a fetch, a render and a sandbox, and the model cannot see what is drawn inside it unless the app sends context back with ui/update-model-context.

An MCP App lives inside one conversation. Close the chat and the UI is gone; only the server's data persists. When the HTML itself is the thing you want to keep, with versions and several people editing it, you want a page outside the chat rather than a frame inside it. That is what a shared board on Intraface is: the artifact, not the view onto one.

Common mistakes

  • Returning an empty content array. Hosts that do not support the extension show nothing. Return the same information as text even when the UI will usually replace it.
  • Loading scripts from a CDN with no CSP declared. The default policy is default-src 'none', so the request is blocked. Either bundle everything into the HTML or list the origins in _meta.ui.csp under connectDomains and resourceDomains.
  • Treating the app as trusted. It came from the server, and the server might be malicious. Hosts render it in a sandbox and may require approval for the tool calls it makes. Design your app assuming the user will be asked.
  • Using the old ui/resourceUri key. The spec marks that form deprecated; _meta.ui.resourceUri is the one to write.

FAQ

Which clients support MCP Apps?
The extension page lists Claude on the web, Claude Desktop, VS Code with GitHub Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam and Archestra.AI. Support is opt-in on both sides: the client declares io.modelcontextprotocol/ui in its capabilities and the server advertises it back.
Are MCP Apps the same as MCP-UI?
MCP-UI is the community project that pioneered sending UI as MCP resources. MCP Apps is the official extension, proposed as SEP-1865 in November 2025 and marked stable in January 2026; its spec credits MCP-UI and OpenAI's Apps SDK as influences. MCP-UI's client library can render MCP Apps.
Can an MCP App read the chat or the page around it?
No. It runs in a sandboxed iframe that cannot reach the parent page's DOM, cookies or storage. Everything it learns arrives through postMessage from the host, and everything it does goes back the same way, where the host can block or ask you first.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.