Remote Control MCP (Electrobun)

A desktop remote-control MCP server built with Electrobun. It captures a user-selected screen and exposes mouse/keyboard control + screenshots as an MCP (Model Context Protocol) server, reachable from anywhere through a WebSocket tunnel — so external agents and tools can drive the machine.

Native input simulation currently targets Linux (via /dev/uinput). See Setup & permissions.

Working on this codebase with an AI agent? Read AGENTS.md first — it documents the architecture and the non-obvious gotchas.

Features

• Screen capture — pick a display via the Web getDisplayMedia API and take JPEG screenshots (90% quality).

• Synthetic mouse — click at absolute pixel coordinates (left/right/middle).

• Synthetic keyboard — type text (with human-like timing) and press named keys with modifiers.

• Remote MCP server — 5 tools (get_system_info, screenshot, click, type_text, press_key) exposed over a WebSocket tunnel with a stable public URL.

• In-app test panel — exercise clicks/typing/keys locally before going remote.

Related MCP server: win-rdp-mcp

Architecture (short version)

Electrobun runs two processes that talk over RPC:

• Bun process (src/bun/) — OS access: /dev/uinput (FFI), xrandr, persistent storage.

• Webview (src/mainview/) — UI, screenshots, and the MCP server + WebSocket tunnel client.

Native actions are forwarded from the webview to Bun via RPC; screenshots are captured in the webview from the active screen-share stream. Full detail and the data-flow diagram live in AGENTS.md.

src/
├── bun/
│   ├── index.ts      # RPC schema + handlers, window, screen-size detection, KV store
│   └── uinput.ts     # FFI → /dev/uinput: VirtualMouse + VirtualKeyboard
└── mainview/
    ├── index.html    # UI
    ├── index.css
    ├── index.ts      # ScreenCaptureApp + McpPanel
    └── mcp/          # tunnel client + MCP server bridge + tool definitions

Getting started

bun install
bun start            # run (electrobun dev)
bun dev              # run with file watching
bun run build:canary # build / verify it compiles

In the app, click Select Screen & Start Remote, choose a display, and the Remote MCP panel will show the public URL once the tunnel connects.

Setup & permissions

Input simulation creates kernel-level virtual devices through /dev/uinput, which is root-only by default. On startup the app offers to grant access via pkexec (graphical password prompt), running:

chmod 666 /dev/uinput

This is session-only (does not survive reboot). Manual equivalent:

sudo chmod 666 /dev/uinput

Do not run the whole app with sudo.

Remote control via MCP

The app exposes a Streamable HTTP MCP endpoint at the public tunnel URL. Tools:

Smoke test

smoke-mcp.sh drives a demo sequence (click → type "chrome" → down → enter):

./smoke-mcp.sh 'https://layerz.me:4433/mcp/<session>'
# or
MCP_URL='https://layerz.me:4433/mcp/<session>' ./smoke-mcp.sh

Paste the URL from the app's Remote MCP panel.

Security considerations

• Screen sharing requires explicit user consent; the tunnel only connects after consent.

• Once connected, the public URL lets remote clients control mouse/keyboard and read the screen — treat it like a remote-access credential and disconnect when not in use.

• The tunnel never autostarts; it is tied to the screen-selection action.

Notes

• US QWERTY keyboard layout assumed for synthetic typing.

• Tunnel server: wss://layerz.me:4433/connect (already deployed; configurable in tunnel.ts).