vnc_command takes an action and its target — a coordinate, a drag destination, a key or combination, text, a scroll direction and amount — and alongside them a set of millisecond controls: click hold, inter-key delay, drag press, drag step, paste settle, scroll tick. That is the difference between a remote desktop that swallows your input and one that accepts it, and it is tunable per call rather than per install. action_queue runs a list of those actions in one round trip, and task_complete and task_failed end the run with a summary or a reason instead of leaving the caller to infer how it went.
A VNC input driver behind four tools: a single action command, a queue that runs a list of them, and two terminators that report the task done or failed.
- One action per call over VNC: pointer position, drag origin and destination, scroll direction and amount, keys, key combinations and typed text
- Timing set per step in milliseconds — how long a click is held, the gap between keystrokes, how long a drag presses before it moves and how long it settles after
- A list of actions queued and executed in a single call
- An explicit end to the run, with a summary when it worked and a reason when it did not
A reachable VNC server — VNC_HOST, VNC_PORT, VNC_USERNAME and VNC_PASSWORD — and the daemon it drives, given by CLAUDE_KVM_DAEMON_PATH with CLAUDE_KVM_DAEMON_PARAMETERS.
One command plus a key — npx -y claude-kvm, then supply credentials