Web vs Desktop Boundary

Capabilities requiring a trust boundary

Provider OAuth, quota routing, model execution, MCP installation and instruction sync, filesystem and shell access, browser automation, secret storage, and scheduled jobs. None of these can be safely implied by a static page.

How stubs behave

A stub page names the capability, its next real boundary, and refuses to simulate connected output. No fabricated results, no fake "processing" state, no implied success.

When a stub becomes live

Only after the desktop shell or a trusted service implements the capability behind an explicit user grant. The boundary flips on connection, never on a page promise.