docs/apps/low-level.mdx
import { VersionBadge } from '/snippets/version-badge.mdx'
<VersionBadge version="3.0.0" />Everything on this page is for when you want full control: your own HTML, your own JavaScript framework, a map library, a 3D viewer, custom video playback. Interactive Tools wrap the MCP Apps extension so you never have to think about it — this page is what you reach for when you need to think about it.
You'll be working with two things: the @modelcontextprotocol/ext-apps JavaScript SDK for host communication, and FastMCP's AppConfig for resources and CSP.
An MCP App has two parts:
ui:// resource containing the HTML that renders that dataThe tool declares which resource to use via AppConfig. When the host calls the tool, it also fetches the linked resource, renders it in a sandboxed iframe, and pushes the tool result into the app via postMessage. The app can also call tools back, enabling interactive workflows.
import json
from fastmcp import FastMCP
from fastmcp.apps import AppConfig, ResourceCSP
mcp = FastMCP("My App Server")
# The tool does the work
@mcp.tool(app=AppConfig(resource_uri="ui://my-app/view.html"))
def generate_chart(data: list[float]) -> str:
return json.dumps({"values": data})
# The resource provides the UI
@mcp.resource("ui://my-app/view.html")
def chart_view() -> str:
return "<html>...</html>"
AppConfig controls how a tool or resource participates in the Apps extension. Import it from fastmcp.server.apps:
from fastmcp.apps import AppConfig
On tools, you'll typically set resource_uri to point to the UI resource:
@mcp.tool(app=AppConfig(resource_uri="ui://my-app/view.html"))
def my_tool() -> str:
return "result"
You can also pass a raw dict with camelCase keys, matching the wire format:
@mcp.tool(app={"resourceUri": "ui://my-app/view.html"})
def my_tool() -> str:
return "result"
The visibility field controls where a tool appears:
["model"] — visible to the LLM (the default behavior)["app"] — only callable from within the app UI, hidden from the LLM["model", "app"] — bothThis is useful when you have tools that only make sense as part of the app's interactive flow, not as standalone LLM actions.
@mcp.tool(
app=AppConfig(
resource_uri="ui://my-app/view.html",
visibility=["app"],
)
)
def refresh_data() -> str:
"""Only callable from the app UI, not by the LLM."""
return fetch_latest()
| Field | Type | Description |
|---|---|---|
resource_uri | str | URI of the UI resource. Tools only. |
visibility | list[str] | Where the tool appears: "model", "app", or both. Tools only. |
csp | ResourceCSP | Content Security Policy for the iframe. |
permissions | ResourcePermissions | Iframe sandbox permissions. |
domain | str | Stable sandbox origin for the iframe. |
prefers_border | bool | Whether the UI prefers a visible border. |
Resources using the ui:// scheme are automatically served with the MIME type text/html;profile=mcp-app. No need to set it manually.
@mcp.resource("ui://my-app/view.html")
def my_view() -> str:
return "<html>...</html>"
The HTML can be anything — a full single-page app, a simple display, or a complex interactive tool. The host renders it in a sandboxed iframe and establishes a postMessage channel for communication.
Your HTML app communicates with the host using the @modelcontextprotocol/ext-apps JavaScript SDK. The simplest approach is to load it from a CDN:
<script type="module">
import { App } from "https://unpkg.com/@modelcontextprotocol/[email protected]/app-with-deps";
const app = new App({ name: "My App", version: "1.0.0" });
// Receive tool results pushed by the host
app.ontoolresult = ({ content }) => {
const text = content?.find(c => c.type === 'text');
if (text) {
document.getElementById('output').textContent = text.text;
}
};
// Connect to the host
await app.connect();
</script>
The App object provides:
app.ontoolresult — callback that receives tool results pushed by the hostapp.callServerTool({name, arguments}) — call a tool on the server from within the appapp.onhostcontextchanged — callback for host context changes (e.g., safe area insets)app.getHostContext() — get current host contextSee the full ext-apps SDK documentation for the complete API reference.
<Note> If your HTML loads external scripts, styles, or makes API calls, you need to declare those domains in the CSP configuration. See [Security](#security) below. </Note>Apps run in sandboxed iframes with a deny-by-default Content Security Policy. By default, only inline scripts and styles are allowed — no external network access.
If your app needs to load external resources (CDN scripts, API calls, embedded iframes), declare the allowed domains with ResourceCSP:
from fastmcp.apps import AppConfig, ResourceCSP
@mcp.resource(
"ui://my-app/view.html",
app=AppConfig(
csp=ResourceCSP(
resource_domains=["https://unpkg.com", "https://cdn.example.com"],
connect_domains=["https://api.example.com"],
)
),
)
def my_view() -> str:
return "<html>...</html>"
| CSP Field | Controls |
|---|---|
connect_domains | fetch, XHR, WebSocket (connect-src) |
resource_domains | Scripts, images, styles, fonts (script-src, etc.) |
frame_domains | Nested iframes (frame-src) |
base_uri_domains | Document base URI (base-uri) |
If your app needs browser capabilities like camera or clipboard access, request them via ResourcePermissions:
from fastmcp.apps import AppConfig, ResourcePermissions
@mcp.resource(
"ui://my-app/view.html",
app=AppConfig(
permissions=ResourcePermissions(
camera={},
clipboard_write={},
)
),
)
def my_view() -> str:
return "<html>...</html>"
Hosts may or may not grant these permissions. Your app should use JavaScript feature detection as a fallback.
This example creates a tool that generates QR codes and an app that renders them as images. It's based on the official MCP Apps example. Requires the qrcode[pil] package.
import base64
import io
import qrcode
from mcp import types
from fastmcp import FastMCP
from fastmcp.apps import AppConfig, ResourceCSP
from fastmcp.tools import ToolResult
mcp = FastMCP("QR Code Server")
VIEW_URI = "ui://qr-server/view.html"
@mcp.tool(app=AppConfig(resource_uri=VIEW_URI))
def generate_qr(text: str = "https://gofastmcp.com") -> ToolResult:
"""Generate a QR code from text."""
qr = qrcode.QRCode(version=1, box_size=10, border=4)
qr.add_data(text)
qr.make(fit=True)
img = qr.make_image()
buffer = io.BytesIO()
img.save(buffer, format="PNG")
b64 = base64.b64encode(buffer.getvalue()).decode()
return ToolResult(
content=[types.ImageContent(type="image", data=b64, mimeType="image/png")]
)
@mcp.resource(
VIEW_URI,
app=AppConfig(csp=ResourceCSP(resource_domains=["https://unpkg.com"])),
)
def view() -> str:
"""Interactive QR code viewer."""
return """\
<!DOCTYPE html>
<html>
<head>
<meta name="color-scheme" content="light dark">
<style>
body { display: flex; justify-content: center;
align-items: center; height: 340px; width: 340px;
margin: 0; background: transparent; }
img { width: 300px; height: 300px; border-radius: 8px;
box-shadow: 0 2px 8px rgba(0,0,0,0.1); }
</style>
</head>
<body>
<div id="qr"></div>
<script type="module">
import { App } from
"https://unpkg.com/@modelcontextprotocol/[email protected]/app-with-deps";
const app = new App({ name: "QR View", version: "1.0.0" });
app.ontoolresult = ({ content }) => {
const img = content?.find(c => c.type === 'image');
if (img) {
const el = document.createElement('img');
el.src = `data:${img.mimeType};base64,${img.data}`;
el.alt = "QR Code";
document.getElementById('qr').replaceChildren(el);
}
};
await app.connect();
</script>
</body>
</html>"""
The tool generates a QR code as a base64 PNG. The resource loads the MCP Apps JS SDK from unpkg (declared in the CSP), listens for tool results, and renders the image. The host wires them together — when the LLM calls generate_qr, the QR code appears in an interactive frame inside the conversation.
Not all hosts support the Apps extension. You can check at runtime using the tool's context:
from fastmcp import Context
from fastmcp.apps import AppConfig, UI_EXTENSION_ID
@mcp.tool(app=AppConfig(resource_uri="ui://my-app/view.html"))
async def my_tool(ctx: Context) -> str:
if ctx.client_supports_extension(UI_EXTENSION_ID):
# Return data optimized for UI rendering
return rich_response()
else:
# Fall back to plain text
return plain_text_response()