# One-line install

Installs OpenClaw, the skills and a model configuration in a single command,
without a USB drive.

## Install

**Mac / Linux**

```bash
curl -fsSL https://u-claw.org/install.sh | bash
```

**Windows** (PowerShell)

```powershell
irm https://u-claw.org/install.ps1 | iex
```

**From a checkout**

```bash
bash install/install.sh                                          # Mac / Linux
powershell -ExecutionPolicy Bypass -File install\install.ps1     # Windows
```

> These pipe a script from the network straight into a shell. That pattern is
> widely warned about, and on a managed corporate machine an EDR agent will
> often block it outright. If you are on a work laptop, use the USB build
> instead — that is what it is for.

## What it does

| Step | Mac / Linux | Windows |
|---|---|---|
| 1 | Node.js — into `~/.uclaw/runtime/`, or reuses the system copy if it is v20+ | same, under `%USERPROFILE%\.uclaw\` |
| 2 | OpenClaw via npm | downloads a pre-built bundle, **checksum verified** |
| 3 | QQ plugin — optional, a failure here does not stop anything else | same |
| 4 | Skills — read from `skills/manifest.json`, fetched from the repo | same |
| 5 | Model — pick one and paste a key | same |
| 6 | Start script — `start.command` / `start.sh` | `start.bat` |
| 7 | Verify what landed | same |

Neither script contains any skill text of its own. Both call
`lib/install-skills.mjs`, which is the same installer the USB build and the
release pipeline use — that is what keeps the two platforms from drifting apart,
which they had badly done before.

## Models

The menu leads with what is easiest to get working:

| | Model | Why |
|---|---|---|
| 1 | **Gemini** | Easiest key to obtain — Google account, no card, free tier |
| 2 | **Claude** | Best all-rounder |
| 3 | **GPT** | Most widely used |
| 4 | **OpenRouter** | One key across many models |
| 5 | **SEA-LION** | Malay, Tamil and Singlish. Free, but rate-limited to 10 requests a minute |
| 6 | **Local** | Runs offline through Ollama |
| 7 | Other | Paste your own endpoint and model name |

Others — DeepSeek, Kimi, Qwen, GLM and so on — work fine through option 7.

## Where things land

```
~/.uclaw/
├── runtime/node-{platform}/       Node.js
├── core/                          OpenClaw and plugins
│   └── node_modules/openclaw/skills/     installed skills
├── data/
│   ├── .openclaw/openclaw.json    your model and key
│   ├── memory/                    what the AI remembers
│   └── backups/
└── start.command | start.sh | start.bat
```

## Notes

- **Nothing touches your system Node.** It installs into `~/.uclaw/runtime/`, and
  reuses a system copy only if it is already v20 or newer.
- **Download sources** are `nodejs.org` and `registry.npmjs.org`; the Windows
  bundle comes from our own releases and its SHA256 is checked before it is
  unpacked. Until recently the only check was "is the file bigger than 1MB",
  which is not a check.
- **Piped mode**: under `curl | bash` there is no terminal to prompt on, so it
  takes the defaults and you set the model afterwards.
- **Uninstall**: `rm -rf ~/.uclaw`, or delete `%USERPROFILE%\.uclaw` on Windows.

## How this relates to the other builds

```
install/     one-line install over the network — this module
portable/    the USB build, the main product
bootable/    a Linux USB that boots a machine with no OS
u-claw-app/  Electron desktop app — deprecated, kept for archive
```

The result is the same as running `advanced/Mac-Install.command` from the USB
build. The difference is only that everything is fetched from the network rather
than copied off a drive.