Desktop commands

The workspace image ships a desktop: KasmVNC as the X server and browser viewer, Openbox, Firefox, and x11vnc on the same display so kobe vnc can still speak RFB. An agent can look at the screen and act without a browser.

On your laptop, kobe vnc open starts the desktop, forwards KasmVNC on 6080, and opens it in the browser. --native is TigerVNC against 5900 if vncviewer is on PATH. macOS Screen Sharing is not used: it demands a VNC password, and x11vnc has none so screenshot/click can speak RFB security None.

kobe vnc open

kobe vnc open dev

Starting the desktop in dev...
Opened http://127.0.0.1:52341/
Ctrl-C closes the viewer; the sandbox keeps running.

The forward runs until you interrupt it. Interrupting closes the browser view, not the desktop: the sandbox keeps running and kobe vnc open reconnects to the same session.

--no-browser prints the URL instead of opening one. --local-port pins the port. --no-start skips starting the desktop.

Starting the desktop by hand

A lease does not start a desktop, and kobe vnc open is doing this for you:

kobe exec dev --detach -- kobe-desktop start

It runs until the lease ends or you cancel the execution. Only one desktop runs in a Sandbox at a time. kobe-desktop status, stop, logs, terminal and browser are also available; kobe-desktop browser <url> opens Firefox inside the desktop, which is different from kobe vnc open opening the desktop in the browser on your own machine.

kobe vnc screenshot

Write the desktop to a PNG file on your machine.

kobe vnc screenshot dev -f desktop.png

Wrote desktop.png (1440x900)

The image is read over the connection and written locally. It never passes through the execution API, which matters: kobe exec converts its output to text, so capturing to a file inside the Sandbox and printing it back would corrupt it.

kobe vnc click

kobe vnc click dev --at 400,300
kobe vnc click dev --at 400,300 --button right

Coordinates are pixels from the top-left corner. A point outside the desktop is refused rather than sent and silently ignored.

Use kobe vnc move to move the pointer without pressing anything, which is how you trigger a hover.

kobe vnc type, kobe vnc key, and kobe vnc paste

kobe vnc type dev --text 'echo hello > /tmp/note'
kobe vnc key dev --key return
kobe vnc paste dev --text 'https://example.com/login'

paste puts the text on the X clipboard. Click the field in the desktop and press Ctrl+V (Linux, even when the viewer is on a Mac). That is how a URL or password gets into Firefox when you cannot paste from the host.

type sends printable ASCII, holding Shift where the character needs it. key presses one named key: return, tab, escape, backspace, delete, home, end, pageup, pagedown, space, and the four arrows.

A character outside printable ASCII is refused by name rather than typed as something else. Accents and international layouts need a layout-aware keysym table that does not exist yet, and a silently wrong character is worse than an error that says so.

The loop

Screenshot, decide, act, screenshot again:

kobe vnc screenshot dev -f before.png
kobe vnc click dev --at 400,300
kobe vnc type dev --text 'make test'
kobe vnc key dev --key return
kobe vnc screenshot dev -f after.png

Input alone is not enough, because something has to decide where to click. That is why both halves ship together.

What it does not need

No port has to be open and no browser has to be running. The connection is one short-lived loopback forward carried by the same authenticated stream as kobe port-forward, so it inherits the same authorization and costs nothing against the execution budget.

kobe port-forward dev 6080:6080 still works if you would rather wire it up yourself. kobe vnc open is that plus starting the desktop. --native plus TigerVNC is the RFB path.

Limits

• The VNC server must be the one the workspace image starts, listening on loopback inside the Sandbox. A server asking for a VNC password is refused: kobe reaches the desktop over an already-authenticated stream and holds no VNC credentials. The exception is --host, which logs in to a Mac's Screen Sharing with a password you configure; see host commands.
• Colour-map servers are refused rather than rendered with the wrong colours.
• A framebuffer larger than 64 megapixels is refused rather than allocated.