BrowserPreview

Browser

Every workspace keeps separate human and automation browser surfaces. Developers use Desktop or Mobile views. Agents use the Server headless profile.

Those surfaces do not share cookies, storage, or navigation state. Restarting applications should reload the existing surface rather than create more tabs.

1. Start the workspace so HTTP ports exist

2. Open the human surface or run an MCP browser action

3. Treat the two profiles as independent browsers

Rule

Profiles never share storage

Where it differs

Desktop

Embedded views connect directly to each application's allocated HTTP port.

Mobile

The application display uses an authenticated HTTPS reverse proxy on the Server. Tickets travel in a request header or URL fragment, never as a query string.

MCP

MCP browser tools use the Server headless Chromium profile. They do not read Desktop or Mobile storage.

Surfaces

Surface	Owner	Storage	Used by
Desktop embedded browser	Desktop	Native view runtime on the workstation	Humans using HTTP port tabs in Agent-Up Desktop
Mobile application view	Server proxy + client view	Independent of Desktop and headless	Humans on Android, iOS, and the PWA
Headless automation browser	Server	browser-profiles/{workspaceId}	MCP browser tools

These profiles do not share cookies, local storage, session storage, IndexedDB, cache, or navigation state. A login, cookie, or page visit in Desktop is not visible to the headless session, and agent actions in the headless session do not update Desktop.

Session continuity

Within a workspace, Desktop preserves each HTTP port tab's view state when you switch tabs or applications. The Server preserves the headless profile across MCP browser actions and application restarts.

Changing workspaces restores the selected workspace's saved state on each surface independently. Restarting applications should reload the existing Desktop view and headless session rather than opening duplicate tabs or profiles.

Structured inspection

Agent-Up exposes browser state to agents through structured inspection instead of requiring users or agents to scrape raw page markup.

Inspection can include:

• Accessibility tree.
• Interactive elements.
• Page metadata.
• DOM snapshot.
• HTML.
• Browser history.
• Screenshot.

Accessibility data should be preferred over raw HTML.

Next in this slice

Validation

Record and replay a user-meaningful route.