Skip to main content

cua-localhost

Drive the user's host machine directly through Cua, with no sandbox. Maps to Cua's Localhost API. Higher risk - the agent can click, type, and run shell commands on the real desktop. Use only when the user explicitly opts in and read the safety section first. Triggers - "control my computer", "click here", "use my real Chrome", "no sandbox", "내 컴퓨터 자동화".

Datos de origen

Repositorio
code-yeongyu/pi-cua-integration
Última actividad en el origen
13 de mayo de 2026 a las 11:25
Idioma detectado de SKILL.md
inglés
Estrellas
7
Forks
5

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
cua-localhost
description
Drive the user's host machine directly through Cua, with no sandbox. Maps to Cua's Localhost API. Higher risk - the agent can click, type, and run shell commands on the real desktop. Use only when the user explicitly opts in and read the safety section first. Triggers - "control my computer", "click here", "use my real Chrome", "no sandbox", "내 컴퓨터 자동화".
# Localhost (unsandboxed) When `mode: "localhost"` is set in `.pi/cua.jsonc`, Cua connects to your host machine directly. There is no sandbox: every click, keystroke, scroll, and shell command happens on the real OS. ## Activation ```jsonc // .pi/cua.jsonc { "mode": "localhost", "localhost": { "confirmDestructive": true } } ``` ## Tools that work in this mode `cua_screenshot`, `cua_click`, `cua_type`, `cua_key`, `cua_scroll` - all without a `sandbox` argument. For host shell execution use Pi's built-in `bash` tool. ComputerAgent-style autonomous delegation is no longer shipped here - the main pi agent loop drives the cua surface directly, and bash-level `cua do task ...` invocations are documented in the global `cua-skill` for callers that explicitly want sub-agent delegation. `cua_sandbox_start`, `cua_sandbox_stop`, `cua_sandbox_list` are disabled. ## Required permissions ### macOS - **Accessibility** - System Settings > Privacy & Security > Accessibility > add your terminal / Pi binary - **Screen Recording** - same panel, Screen Recording - **Automation** - approve when prompted, or add manually for the apps you want to drive ### Linux - X11 desktop (Wayland is NOT supported by Cua's localhost mode) - `cua-auto` controls input through pynput and captures screenshots through its Python screen backend ### Windows - Localhost mode is best-effort on Windows; prefer local QEMU VM instead. ## Safety guardrails 1. **Confirm destructive shell.** With `confirmDestructive: true` (default) the extension prompts before running shell commands containing patterns like `rm -rf`, `dd`, or `curl ... | sh`. 2. **Never type secrets.** Do not paste API keys or passwords; the model can see what you type. 3. **Prefer sandbox mode** for anything that opens unknown URLs or runs untrusted code. 4. **Stop immediately** if a screenshot shows unexpected content - the agent may have been redirected. ## When to choose localhost - Driving your real browser/IDE because the workflow needs your real cookies/sessions. - Cross-window automation that does not fit in a sandbox. - Lightweight scripting where spinning up a VM is overkill. For anything else, prefer `local` sandbox mode.
Ver en GitHub