w3bs fleet desktop: headless Kubuntu desktop in Docker for AI agents (CLI 0.1.7)

[ralyodio]

On a server with no desktop, one command gives it one and shares it on w3bs.org/fleet:

w3bs fleet desktop start            # view only
w3bs fleet desktop start --control  # also mouse and keyboard from /fleet
w3bs fleet desktop agent claude     # prints the claude mcp add line for the desktop

What runs: a Kubuntu (ubuntu:26.04) container with KDE Plasma, Konsole, the Agentic Desktop Layer (Enabled, MCP 127.0.0.1:7802) and TigerVNC's w0vncserver, which is built from source in the image because Ubuntu's TigerVNC 1.15 has no Wayland server.

No GPU needed: KWin drives the kernel's vkms virtual display, and Mesa composites with OpenGL on the CPU. start loads vkms once through sudo and persists it. The two alternatives I tried both fell back to QPainter, which can't screencast:

• KWin nested inside Xvnc;
• KWin on a vgem render node.

VNC: only on a 0600 Unix socket. There's no TCP port and no password. The agent's --vnc now accepts unix:/path.

Known compromise: KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 inside the container. Ubuntu's KWin 6.6 doesn't trust the KDE portal there, so the screencast had no source. Everything in that compositor belongs to the one agent desktop.

Proof:

• On dev3 (Ubuntu 26.04 VM, no GPU), w3bs fleet desktop start --control took 12 s with the image cached. Production shows dev3 shared, reachable and listening.
• The desktop layer typed into Konsole, and the VNC socket showed Plasma with that text.
• desktop/tests/rig/fleet-screen.sh passes 11/11 over TCP and 11/11 with VNC_UNIX=1.
• CLI tests pass, including 3 new ones. Server tests pass 32/32.

🤖 Generated with Claude Code