AutoBashCMD

English | 中文

A zsh plugin that turns natural language into shell commands using local AI (Ollama / LM Studio).

Type what you want in plain language — get an executable command back. No cloud, no API keys, everything runs locally.

Requirements

• zsh, curl, jq
• Ollama or LM Studio running locally
• Recommended model: qwen2.5-coder:3b — fast, lightweight, good at shell command generation

  ollama pull qwen2.5-coder:3b

Installation

Agent Install: Give this URL to your coding agent and it will handle everything automatically:

https://raw.githubusercontent.com/chaodongzhang/AutoBashCMD/main/INSTALL_GUIDE.md

Manual Install:

git clone https://github.com/chaodongzhang/AutoBashCMD.git ~/.AutoBashCMD

Add to your ~/.zshrc:

source ~/.AutoBashCMD/shell/autobashcmd.zsh

Then reload:

source ~/.zshrc

Usage

ai — Generate a command

ai list all files in the current directory
ai find the process using port 3000
ai compress this folder into a zip

Outputs the command only. Does not execute.

airun — Generate and confirm before executing

airun delete all .log files

Shows the command first, press Enter to run or any other key to cancel. Dangerous commands (e.g. rm, sudo) require typing yes.

aiexplain — Explain a command

aiexplain grep -rn
aiexplain what does chmod 755 mean

Natural language in shell

Just type naturally — if the input looks like natural language (contains non-ASCII characters), it auto-routes through airun:

列出当前目录下的文件
进入 src 目录
找出占用 3000 端口的进程

Help queries also work directly:

ls这个命令如何使用
介绍一下 grep

Configuration

Variable	Default	Description
SHELL_AI_BACKEND	ollama	ollama or lmstudio
SHELL_AI_MODEL	qwen2.5-coder:3b	Model for command generation
SHELL_AI_TEXT_MODEL	same as MODEL	Model for help/explain
SHELL_AI_DEBUG	unset	Set to 1 for debug output
OLLAMA_BASE_URL	http://localhost:11434	Ollama API endpoint
LMSTUDIO_BASE_URL	http://127.0.0.1:1234/v1	LM Studio API endpoint

Switch backends

# Use LM Studio
export SHELL_AI_BACKEND=lmstudio
export SHELL_AI_MODEL=qwen3.5-9b-mlx

# Back to Ollama
export SHELL_AI_BACKEND=ollama
export SHELL_AI_MODEL=qwen2.5-coder:3b

How it works

1. Local fast-path: Common requests (directory listing, CPU usage, etc.) map directly to commands — no AI call needed.
2. Intent classification: The model classifies input as command, cd, help, or reject.
3. Dispatch: Each intent routes to the appropriate handler — command generation, directory resolution, help text, or rejection.
4. Validation: Generated commands are syntax-checked (zsh -n) and verified (whence). Invalid output triggers one retry with a correction prompt.
5. Safety: Risky commands (rm, sudo, dd, etc.) require explicit confirmation.

License

MIT