soffits/dsh-desktop-view

Watch, live from the Web UI, the machine your DeepSeek Harness agent is running on: hear and see what it does, and step in whenever assistance is required (a QR-code login, a CAPTCHA, an unexpected dialog). While you hold control, two-way clipboard access plus file upload and download are available as well.

The desktop stream that cua-spacesd (Cua's in-guest daemon) generates, together with Cua's web viewer, is carried over DSH's existing auth and web server. Nothing extra is exposed: no port or subdomain is opened, and no daemon credential ever reaches the browser.

What you get

• A Desktop tab on the right sidebar, reachable either through the sidebar guide or via the monitor button in the session header.
• A full-screen page at /desktop-view/, served from your DSH origin. It adapts to a phone screen, supports touch, and supports direct linking. Adding ?control=1 makes it open directly in control mode, letting an agent that needs help send you a link; an agent asking for help sends /desktop-view/?control=1&help=1&session=<its session id>, so that giving control back wakes it to carry on.
• Watching is the default mode. You see and hear the desktop, while clicks, keystrokes, clipboard and files are refused. Picking "Take control" switches modes; "Give back" returns you to watching. Control likewise ends on its own after a stretch with no input, or when the view is closed. Merely hiding the Desktop tab (another sidebar tab, a collapsed sidebar, another session) keeps your control; the stream pauses until you look again.
• All the desktop viewer's features while you hold control: mouse, keyboard and touch input, an on-screen keyboard, clipboard syncing, drag-and-drop file uploads, file browsing and downloads, display selection, scaling, and the microphone wherever the desktop supports one.
• The agent is kept informed. For as long as you hold control, a small JSON file is written by the plugin; the agent ought to read it before acting on the desktop, and each control change notifies the session that opened the view.

Requirements

• DSH 0.2.2-rc.1 or later.
• cua-spacesd 0.12 or later on the same machine, with the DSH process able to reach the env token (on Cua's desktop image that means 127.0.0.1:3211, the token being at /run/cua/env-token).
• An up-to-date browser: Chrome, Edge, Firefox or Safari 17+. On mobile, both Safari for iOS and Chrome for Android are supported.

Installation

Directly from GitHub:

dsh plugin --profile web add github:soffits/dsh-desktop-view

Alternatively, add the bundle through the Web UI's plugin manager. Restart DSH afterwards.

Configuration

Every option appears as a field in the desktop-view plugin row.

Setting	Default	Description
endpoint	http://127.0.0.1:3211	Address on which cua-spacesd listens. Leave it on loopback, as this address receives the env token.
tokenFile	/run/cua/env-token	File containing the env token; it stays on the server at all times.
maxFps	30	Frame rate requested by the viewer. Reducing it lowers CPU and bandwidth use.
clipboard	true	Clipboard syncing while you hold control.
filesRoot	~	Desktop folder you can browse, upload to and download from while you hold control. Leaving it empty disables file transfer.
microphone	true	Permit sending your microphone to the desktop while you hold control, where the desktop supports it.
idleSeconds	120	After this many seconds with no input, control reverts to watching.
stateFile	/tmp/dsh-desktop-view/control.json	Path at which the plugin writes the control state that the agent reads.
notifyAgent	wake	Notify the session that opened the view whenever control changes: wake (returning control wakes the agent if it is running or sent the help link; otherwise the note waits), quiet (notes always wait for the agent's next step), off.

What the agent can see

As an example, stateFile contains:

{
  "controller": "user",
  "since": "2026-11-08T12:00:00.000Z",
  "sessionId": "01H…",
  "updatedAt": "2026-11-08T12:00:04.120Z"
}

When the owner holds control, controller is "user"; otherwise it is null. In the null case, the file also records the most recent reason control ended (lastReason: released, idle, disconnected, replaced, stopped). Any script or skill that drives the desktop should check it first and yield whenever the value is not null.

When notifyAgent is on, the session that opened a view gets a short note each time someone takes control (queued; the agent is never woken for it) and each time control returns. The return note wakes the agent only when it has to carry on: it is still running, or the view came from its help link (help=1). Otherwise it waits quietly for the agent's next step, so no turn starts for nothing. In the chat the notes appear as a small "Desktop control" row.

Security

• All requests aimed at the plugin's paths pass through DSH's admission checks: the Host/Origin fence first, blocking cross-site and DNS-rebinding requests, then the browser's signed cookie, and no principal but the owner's is let in.
• The env token of cua-spacesd, along with the viewer tickets minted from it by the plugin, never leave the Host process. The browser keeps only short-lived relay tickets, which function through this relay together with your DSH sign-in, and only for as long as the view (and, for control, your control) persists.
• The server decides what the browser is allowed to do. Watching uses a view-only viewer ticket, so input, clipboard and files are refused by the daemon itself; beyond that, only the media messages and calls that the viewer requires in each mode are forwarded by the relay. Only one view holds control at a time, and control ends once that view disappears.
• The plugin's own pages go out with no-store, no referrer, and a content security policy that is strict and same-origin.

Licensing

MIT. The desktop viewer displayed is Cua's web viewer, which your own cua-spacesd serves at runtime under Cua's FSL-1.1-MIT licence; the plugin here is independent of Cua and neither includes nor redistributes that viewer.