cua

Files (1)

• SKILL.md

SKILL.md (first lines)

# CUA — Computer Use Agent

Autonomous browser control powered by three cloud browser providers: **Steel**, **Browser Use**, and **Browserbase**. Each offers CDP-based headful browser sessions for Playwright/Puppeteer automation, LLM-driven visual browsing, and human-in-the-loop workflows.

## Quick Start

```bash
# Environment setup
export STEEL_API_KEY="ste-..."
export BROWSERUSE_API_KEY="bu_..."
export BROWSERBASE_API_KEY="bb_live_..."
export BROWSERBASE_PROJECT_ID="446f4b4c-..."

# Daemon commands
/cua new                          # Create a Steel session (default provider)
/cua new --provider browserbase   # Create a Browserbase session
/cua new --provider browseruse    # Create a Browser Use session
/cua browse https://example.com   # Navigate in active session
/cua screenshot                   # Capture current page
/cua status                       # Show session details + live view URL
/cua release                      # End session
/cua list                         # List all active sessions
```

## Architecture

```
┌──────────────────────────────────────────────────────────────────┐
│                      SolanaOS CUA Skill                          │
│                                                                   │
│  ┌──────────────┐  ┌───────────────┐  ┌───────────────────────┐ │
│  │  Steel        │  │  Browser Use   │  │  Browserbase          │ │
│  │  Provider     │  │  Provider      │  │  Provider             │ │
│  │              │  │               │  │                       │ │
│  │  • Sessions   │  │  • CDP URL     │  │  • Sessions           │ │
│  │  • Proxy      │  │  • Cloud       │  │  • Extensions         │ │
│  │  • CAPTCHA    │  │  • Profiles    │  │  • Stealth Mode       │ │
│  │  • Live View  │  │  • Proxy       │  │  • Proxies            │ │
│  │  • Recording  │  │               │  │  • Contexts           │ │
│  │  • Mobile     │  │               │  │  • Functions           │ │
│  │  • Regions    │  │               │  │  • Live View           │ │
│  │  • HLS Replay │  │               │  │  • Session Recording   │ │
│  └──────┬───────┘  └──────┬────────┘  └──────────┬────────────┘ │
│         │                  │                       │              │
│  ┌──────┴──────────────────┴───────────────────────┴───────────┐ │
│  │              CDP (Chrome DevTools Protocol)                  │ │
│  │                    WebSocket Connection                      │ │
│  └──────────────────────────┬──────────────────────────────────┘ │
│                              │                                    │
│  ┌──────────────────────────┴──────────────────────────────────┐ │
│  │           Playwright / Puppeteer / Selenium                  │ │
│  │                                                              │ │
│  │  • Page navigation      • Form automation                   │ │
│  │  • Element interaction   • Screenshot capture                │ │
│  │  • File download/upload  • PDF generation                    │ │
│  │  • Network interception  • Cookie management                 │ │
│  └──────────────────────────────────────────────────────────────┘ │
│                                                                   │
│  ┌──────────────────────────────────────────────────────────────┐ │
│  │              LLM Vision Agent (Optional)                      │ │
│  │                                                              │ │
│  │  Screenshot → Vision LLM → Action Plan → Playwright Exec    │ │
│  │                                                              │ │
│  │  Supported: Claude, GPT-4o, Gemini, Llama 3.2, Qwen 2.5 VL │ │
│  └──────────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘
```

## Provider Comparison

| Feature | Steel | Browser Use | Browserbase |
|---------|-------|-------------|-------------|
| CDP WebSocket | ✅ | ✅ | ✅ |
| Playwright | ✅ | ✅ | ✅ |
| Puppeteer | ✅ | ✅ | ✅ |
| Selenium | ✅ | ❌ | ✅ |
| Live View | ✅ WebRTC | ❌ | ✅ iframe |
| Session Recording | ✅ HLS/MP4 | ❌ | ✅ rrweb |
| Proxy | ✅ Residential | ✅ | ✅ |
| CAPTCHA Solving | ✅ Auto | ❌ | ✅ Auto |
| Stealth Mode | ✅ Fingerprint | ❌ | ✅ Advanced |
| Mobile Mode | ✅ | ❌ | ✅ Viewport |
| Multi-Region | ✅ LAX/ORD/IAD | ❌ | ✅ Multiple |
| Extensions | ❌ | ❌ | ✅ |
| Contexts (persist state) | ❌ | ✅ Profiles | ✅ |
| Keep Alive | ❌ | ❌ | ✅ |
| Functions (serverless) | ❌ | ❌ | ✅ |
| Max Timeout | 24h | N/A | Plan-based |
| Human-in-the-loop | ✅ interactive | ❌ | ✅ interactive |

## Steel Integration

### Creating Sessions

```go
// One-line Playwright connection (simplest)
browser := chromium.ConnectOverCDP("wss://connect.steel.dev?apiKey=YOUR_KEY")

// Or with full options
session := steel.Create(ctx, &steel.CreateOptions{
    UseProxy:     true,
    SolveCaptcha: true,
    Timeout:      600000, // 10 minutes
    Region:       "lax",
    DeviceConfig: &steel.DeviceConfig{Device: "mobile"},
})
```

### Session Lifecycle

```
CREATE → LIVE → [automation] → RELEASE
                     ↓
              FAILED (crash/timeout)
```

- Sessions auto-release after timeout (default 5 min)
- CDP inactivity timeout: 10 min without commands
- Send heartbeat every 5 min to keep alive: `page.evaluate(() => undefined)`

From x402agent/SolanaOS, MIT.

More from x402agent/SolanaOS