Skills that let an AI agent operate a computer through an ESP-KVM - open-source IP-KVM firmware for the ESP32-P4.
A KVM already lets a person see a machine that has no working operating system. This device can also hand back what a BIOS or a boot loader is showing as exact characters rather than as a picture, so a model can read the firmware menu, decide, and press a key.
Everything here talks to the device over its normal HTTPS API. Nothing is installed on the KVM, and nothing here reaches the internet.
/plugin marketplace add espkvm/skills
/plugin install espkvm@espkvm
That gives you /espkvm:operate. Claude also loads it on its own when a task is
clearly about driving a machine through a KVM.
To try it without installing, copy plugins/espkvm/skills/operate/ into
~/.claude/skills/ (then it is /operate) or into a project's
.claude/skills/.
| Skill | |
|---|---|
operate |
sign in, read the screen as text, keyboard and pointer, virtual media, runbooks, power, settings - and the traps that cost an hour if nobody tells you |
The keyboard, pointer and screenshot endpoints are off by default on the device. Turn on Agent REST API in Settings -> Security. It is off because it hands a program the same control over the target that the console has.
Reading the screen as text, runbooks, virtual media and power work without it.
- The firmware itself: espkvm/espkvm
- An MCP server wrapping the same calls as tools, for hosts that speak MCP: espkvm/mcp
The skill and the MCP server are two answers to the same question. The skill
suits an agent that can run curl and wants the whole API; the MCP server suits
a host that wants a fixed set of tools with a permission switch on input and
power.
Apache-2.0, like the firmware.