Security & CSP

Hosts render Views in a sandboxed iframe with a strict Content Security Policy. By default a View can run its own inlined JS and CSS (which is how moonpeak build ships it) and talk to the host over the bridge. It cannot fetch other origins, load remote scripts or open popups on its own.

Allowing more

export const resource = {
  title: "Map",
  csp: {
    connectDomains: ["https://api.example.com"],     // fetch / WebSocket
    resourceDomains: ["https://tiles.example.com"],  // images, fonts, scripts
  },
  permissions: { geolocation: {} },
};

moonpeak sends these as MCP Apps _meta.ui.csp / permissions and as ChatGPT's openai/widgetCSP. Ask only for what you use; hosts may review apps that ask for more.

Good practice

• Prefer useCallTool over fetch: the call goes through the host to your server, with no CSP entry and no secrets in the View.
• Never put API keys in a View. Everything in it is visible to the user.
• Treat tool input as untrusted: validate with schema, and check permissions in handlers, including app-only ones.
• Use useOpenLink for links so the host can show its own confirmation.