Switch Use — AI skill for Claude Code

switch-use

**Let AI agents drive a real Nintendo Switch.** Read the text on screen, tap a label, wait for a screen, press buttons and sticks, type, take screenshots — from the command line or from Claude Code, Codex or any agent that can run a shell command. Computer-use, but for the Switch.

**让 AI agent 操作真实的任天堂 Switch。** 读屏幕文字（OCR）、按文字点击、等待画面出现、按键和摇杆、 触摸、键盘输入、截图。主机跑大气层（Atmosphère）+ sys-botbase，电脑和主机在同一个局域网即可。

• One ~1 MB Rust binary, no daemon, no Python. macOS and Linux.
• Talks to sys-botbase over TCP on your LAN.
• Screen text via macOS Vision OCR (Chinese, Japanese, English) with the point to touch.
• Ships an agent skill (SKILL.md); part of the `*-use` family (iphone-use, chrome-use, mail-use, …).
• Write-up with real-console stories: English · 中文.

Install

curl -fsSL https://raw.githubusercontent.com/leeguooooo/switch-use/main/install.sh | sh

Apple Silicon Macs and Linux (x86_64, arm64) get a prebuilt binary from GitHub Releases. Intel Macs: `cargo install --git https://github.com/leeguooooo/switch-use`. Update later with `switch-use upgrade`.

As an agent skill:

• Claude Code: /plugin marketplace add leeguooooo/plugins, then /plugin install switch-use@leeguooooo-plugins (or the whole family: use-family).
• Codex / other agents: npx skills add leeguooooo/switch-use.

On the Switch (Atmosphère, booted into emuMMC):

1. Copy sys-botbase to the SD card (atmosphere/contents/430000000000000B/) and reboot.
2. Join the same Wi-Fi as the computer. switch-use discover finds it.

What an agent sees

The Switch has no accessibility tree, so `elements` is the stand-in: a screenshot, read with OCR, every line of text with the point to touch.

$ switch-use elements
screens