Stop memorizing commands. Just say what you need.
NL2Shell turns plain English into terminal commands — instantly, locally, privately. No API keys. No cloud dependency. A fine-tuned 0.8B parameter model that runs on a Raspberry Pi.
$ nl2shell "find all Python files larger than 1MB modified this week"> find . -name "*.py" -size +1M -mtime -7Every developer has typed "how to tar.gz a folder" into a search engine. ChatGPT works but requires internet, costs money, and sends your workflow context to someone else's server. Shell commands have cryptic syntax that even experienced developers look up constantly.
We fine-tuned Qwen3.5-0.8B (a hybrid DeltaNet architecture) using QLoRA on 12,834 curated natural-language-to-bash pairs. The result is a model small enough for edge devices but accurate enough for daily use.
Key properties:
- Runs fully offline after a one-time 400MB download
- Sub-second inference on Apple Silicon, ~1s on Raspberry Pi 5
- Covers Linux, macOS, git, Docker, Kubernetes, cloud CLIs
- MIT licensed — use it however you want
| Repo | What | Stack |
|---|---|---|
| nl2shell | Core ML — model training, datasets, benchmarks | Python, PyTorch, Unsloth, QLoRA |
| nl2shell-web | Product website + MCP server at nl2shell.com | Next.js, React, Tailwind, Supabase |
| vox | Voice-powered CLI — talk to your terminal | Python, Ollama, speech recognition |
| sandbox-bash-mcp | Sandboxed command execution via MCP protocol | Node.js, Docker, MCP |
| collab | Visual agent workspace with infinite canvas | Electron, TypeScript, xterm.js |
You (plain English)
|
+----------+----------+
| | |
[web] [vox] [collab]
browser voice desktop
| | |
+-----+----+----------+
|
[nl2shell model]
local via Ollama
|
[sandbox-bash-mcp]
safe execution
|
terminal
nl2shell is the brain — the fine-tuned model that translates natural language to shell commands.
Three interfaces feed into it:
- nl2shell-web — browser-based, includes an MCP server so AI agents can call it programmatically
- vox — speak commands aloud, get shell output. Fully offline via Ollama
- collab — desktop app with infinite canvas for managing multiple terminal sessions visually
sandbox-bash-mcp provides safe execution — commands run in isolated Docker containers with session persistence, so generated commands can be tested before running on your real system.
Option 1: Try it in the browser Visit nl2shell.com
Option 2: Install the CLI (recommended)
curl -fsSL https://nl2shell.com/install.sh | sh
nl2shell "list all running docker containers sorted by memory"Option 3: Run the model directly
ollama run hf.co/AryaYT/nl2shell-0.8b "show disk usage by directory"Option 4: Voice mode
pip install vox-cli
vox "find large log files"NL2Shell exposes three tools via the Model Context Protocol:
| Tool | Purpose |
|---|---|
leshell_translate | Convert natural language to a shell command |
leshell_execute | Run a command in a sandboxed session |
leshell_explain | Explain what a command does |
Point your MCP client at https://nl2shell.com/api/mcp to use them.
See agents.txt and llms.txt for full agent integration specs.
| Property | Value |
|---|---|
| Base model | Qwen3.5-0.8B |
| Parameters | 859M |
| Architecture | Hybrid DeltaNet (25% softmax attention, 75% linear recurrent) |
| Training method | QLoRA (4-bit NF4, rank 16, alpha 32) |
| Training data | 12,834 NL-to-bash pairs (dataset) |
| GGUF sizes | 400MB (q4_k_m) / 650MB (q8_0) |
| License | MIT |
| HuggingFace | AryaYT/nl2shell-0.8b |
| Live demo | HuggingFace Spaces |
We welcome contributions across all repositories. Each repo has its own CONTRIBUTING.md with setup instructions.
Good first issues are labeled good first issue across the org.
Areas where help is most needed:
- Training data: domain-specific command pairs (Kubernetes, AWS, networking)
- Benchmarking: testing on different hardware (especially ARM, Windows WSL)
- Integrations: editor plugins, shell completions, IDE extensions
- Website: nl2shell.com
- Model: HuggingFace
- Dataset: HuggingFace
- Demo: HuggingFace Spaces
- Author: Arya Teja
Built by Arya Teja. Part of the LeSearch AI ecosystem.