Talk to your terminal. Natural language to shell commands, powered by a local AI model.

Stop memorizing flags. Stop Googling syntax. Just say what you want.

vox > find all python files bigger than 1MB
  find . -name "*.py" -size +1M -type f
  Run it? [Y/n/c] y

curl -fsSL https://raw.githubusercontent.com/aryateja2106/vox/main/install.sh | bash

Requirements: Python 3.9+, Ollama for local inference.

curl -fsSL https://raw.githubusercontent.com/aryateja2106/vox/main/uninstall.sh | bash

# Interactive mode — type naturally, get commands
vox

# Single command
vox "show what's running on port 3000"

# Auto-execute (skip confirmation)
vox -x "show free disk space"

# Use a different model
vox --model llama3.2 "list all docker containers"

# Raw command pass-through (in REPL)
vox > !ls -la

• 100% Local — Runs entirely on your machine. Nothing leaves your terminal. No API keys, no cloud, no subscriptions.
• Tiny Model — Powered by a fine-tuned model under 1 billion parameters. Works on any laptop without a GPU.
• One Curl Install — curl | bash and you're running. No Docker, no npm, no accounts.
• Platform Smart — Detects macOS vs Linux and gives you the right command for your system.
• Voice Ready — Built for voice input from day one. Talk to your terminal like you'd talk to a friend.
• Open Source — MIT licensed. Fork it. Train your own model. Make it yours.

You (English) → Vox → Local Model (Ollama) → Shell Command → You decide (run/copy/skip)

1. You type (or speak) what you want in plain English
2. Vox sends it to a local AI model running on your machine via Ollama
3. The model translates your intent to the correct shell command
4. You review the command and choose to run it, copy it, or skip

The model is a fine-tuned Qwen3.5-0.8B trained on 12,000+ natural language to shell command pairs, including expert-curated patterns from senior engineers.

Vox is configured via environment variables:

The terminal is the most powerful interface on every computer. But it's also the most intimidating.

You know you need to find a large file, or check what's running on a port, or restart a Docker container. You know what you want — you just don't remember how. So you Google it. Or ask ChatGPT and pay $20/month for something a sub-1B model can handle locally.

We believe:

1. Shell commands are a solved problem. There are finite commands, finite flags, finite patterns. A small model can learn them all.
2. Privacy matters. Your terminal history shouldn't go to the cloud. Ever.
3. Portability matters. One curl command on any machine — Linux, macOS — and you're productive.
4. The future is voice. Typing commands is fast. Speaking them is faster. The interface should support both.

Vox is a step toward a world where the terminal understands you — not the other way around.

Vox is powered by NL2Shell, a fine-tuned Qwen3.5-0.8B model trained specifically for natural language to shell command translation.

• Parameters: ~859M (runs on CPU, no GPU needed)
• Training data: 12,000+ deduplicated NL→bash pairs including 960+ expert-curated commands
• Format: GGUF (via Ollama) for efficient local inference
• Latency: <500ms on Apple Silicon, <2s on most CPUs

You can also use any other Ollama model: vox --model llama3.2

git clone https://github.com/aryateja2106/vox.git
cd vox
uv sync
uv run python -m vox

MIT