CommandoCommando
Natural-language CLI agent for the Sui ecosystem. One install on Windows, Linux, or macOS - zero config, grounded intelligence.
Commando
One Command. Zero Config. Grounded Intelligence.
Commando (cmdo) is a local-first CLI agent that turns plain-English (or Vietnamese!) intent into safe, validated sui / walrus / site-builder commands. It downloads the upstream Mysten Labs binaries for you, bootstraps the wallet and config files you would otherwise hit as paper-cuts on first run, and uses an LLM grounded in live --help output to plan each command.
Current release: v0.2.4-beta — cross-platform (Windows / Linux / macOS), GitHub-direct binary downloads, no secrets shipped in the npm tarball.
Why Commando
Zero setup
One npm install puts sui, walrus, and site-builder on your PATH and bootstraps every config file they need.
Grounded LLM
The planner is constrained to commands and flags parsed from each binary's --help output, with strict allowlist validation.
Cross-platform
Same one-line install on Windows (PowerShell), Ubuntu, and macOS (Intel + Apple Silicon).
Safe by default
A safety gate blocks destructive intent (rm -rf /, format, dd if=, shutdown, ...) before any process is spawned.
60-second tour
# 1. Install (any OS, Node 20+)
npm install -g sui-commando@beta
# 2. Configure your LLM provider once
cmdo init
# 3. Talk to the Sui ecosystem
cmdo "create new sui address"
cmdo "give me testnet sui from faucet"
cmdo "build my move package"
cmdo "deploy static site in ./dist to walrus-sites" --site-builderThat's the whole loop. No sui client switch --env testnet, no manual walrus get-wal, no chasing the right site-builder deploy flags — Commando plans the command, validates it against the live skill contract, and streams the output back.
How it works
- Router picks the target binary from explicit flags (
--sui,--walrus,--site-builder) or keyword inference. - Skill loader trims
~/.commando/skills/AGENT.md(auto-generated from--help) to only the relevant tool. - LLM planner (OpenAI or OpenRouter) emits structured
{ binary, args }. The output is parsed defensively and the command path is validated against the allowlist; hallucinated subcommands trigger a retry instead of being executed. - Safety gate rejects destructive patterns on every supported OS.
- Execution engine spawns the real binary, streams stdout/stderr live, and matches the stderr tail against well-known failure patterns to print actionable hints (e.g., "missing WAL coins → try
cmdo \"get wal\"").
Read the full design in Architecture.
What ships in v0.2.4-beta
| Area | Status |
|---|---|
| Windows x86_64 | Stable |
| Linux x86_64 | Stable |
| Linux aarch64 | Stable (site-builder skipped — no upstream build) |
| macOS x86_64 / arm64 | Stable |
| Auto-bootstrap Sui wallet | Yes (~/.sui/sui_config/client.yaml) |
| Auto-bootstrap Walrus config | Yes (~/.config/walrus/client_config.yaml, testnet default) |
| LLM providers | OpenAI, OpenRouter |
| Mock planner | CMDO_LLM_MOCK=1 for offline demos |
| Secret-free tarball | Yes — operator credentials read from env vars only |
Where to go next
Quickstart
Install, configure, and run your first prompt in under 5 minutes.
Installation
Cross-platform install details, env vars, and the optional R2 mirror.
Architecture
Components, data flow, and the safety model in depth.
LLM Setup
Pick a provider, set a key, choose a model.
Sui Cookbook
Copy-paste prompts for wallets, faucet, build, publish.
Walrus Sites
Deploy a static site to Walrus Sites end-to-end.
CLI Reference
Every cmdo subcommand and flag.
Troubleshooting
Common errors and the hints Commando prints for them.
How is this guide?