> ## Documentation Index
> Fetch the complete documentation index at: https://noorle.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Computer

> Computer-use tools for full agent control. Screenshots, mouse, and keyboard automation. Agent-only capability.

<Warning>
  Computer is an Agent-only capability. It cannot be attached to MCP gateways due to its sensitive nature. Only agents can access computer control.
</Warning>

Computer capability gives agents full programmatic control of a desktop environment. Useful for complex tasks requiring visual feedback and interactive control.

## Key Features

* **Screenshots** - Capture current screen state
* **Mouse Control** - Move, click, drag operations
* **Keyboard Input** - Type text, press keys
* **Screen Navigation** - Scroll, zoom, multi-window
* **Real-time Feedback** - Visual loop with agent decisions
* **State Tracking** - Remember screen positions

## How to Enable

### For Agents Only

1. **Agents** > Select Agent > **Settings** > **Capabilities**
2. Search for **Computer**
3. Click **Attach**
4. **Save**

<Note>
  Computer cannot be attached to MCP gateways. It's agent-exclusive for security reasons.
</Note>

## Usage Examples

### Take Screenshot

```
"Take a screenshot of the current desktop"
```

Returns PNG image of desktop with dimensions and visible elements.

### Click on Element

```
"Click the blue button in the top-right corner"
```

Agent:

1. Analyzes screenshot
2. Identifies button coordinates
3. Executes click
4. Captures new screenshot

### Fill Form

```
"Fill the login form with username 'alice' and password 'secret'"
```

Agent:

1. Screenshots form
2. Identifies input fields
3. Clicks on username field
4. Types username
5. Clicks on password field
6. Types password
7. Clicks submit button

### Navigate Multi-step Process

```
"Open the application, go to settings, and change the theme to dark"
```

Agent works through steps with visual feedback.

## Screen Coordinates

Agent receives screen coordinates for all elements:

```
Screenshot resolution: 1920x1080
Button position: x=1850, y=20
```

Agent can:

* Click at coordinates
* Drag between points
* Identify text positions
* Calculate relative positions

## Interaction Types

### Mouse Actions

* **click(x, y)** - Single click
* **double\_click(x, y)** - Double click
* **right\_click(x, y)** - Right/context click
* **drag(x1, y1, x2, y2)** - Drag from point to point
* **move(x, y)** - Move cursor without clicking
* **scroll(direction, amount)** - Scroll up/down/left/right

### Keyboard Actions

* **type(text)** - Type text string
* **key(name)** - Press single key (Enter, Tab, Escape, etc.)
* **hotkey(mod, key)** - Keyboard shortcut (Ctrl+C, Cmd+V, etc.)

### Navigation

* **screenshot()** - Capture current screen
* **wait(seconds)** - Wait for page to load
* **maximize()** - Maximize window
* **minimize()** - Minimize window

## Size Tiers

Each Computer instance is a dedicated virtual machine. Choose a size based on your workload:

| Size         | vCPUs | RAM   | Disk   |
| ------------ | ----- | ----- | ------ |
| x2           | 2     | 2 GB  | 40 GB  |
| x4 (default) | 3     | 4 GB  | 80 GB  |
| x8           | 4     | 8 GB  | 160 GB |
| x16          | 8     | 16 GB | 240 GB |
| x32          | 16    | 32 GB | 360 GB |

### Supported Operating Systems

* Ubuntu 24.04, Ubuntu 22.04
* Debian 12, Debian 11

## Configuration

Optional agent specifications:

```json theme={null}
{
  "computer": {
    "shell_enabled": true,
    "browser_enabled": false,
    "browser_max_tabs": 5,
    "browser_max_download_mb": 50,
    "browser_allowed_domains": []
  }
}
```

| Setting                   | Default | Effect                                 |
| ------------------------- | ------- | -------------------------------------- |
| `shell_enabled`           | true    | Enable shell command execution         |
| `browser_enabled`         | false   | Enable browser subsystem               |
| `browser_max_tabs`        | 5       | Maximum concurrent browser tabs        |
| `browser_max_download_mb` | 50      | Maximum download size (MB)             |
| `browser_allowed_domains` | \[]     | Domain allowlist (empty = all allowed) |

## Browser Subsystem

Computer includes an optional **stateful browser** that persists sessions, cookies, and navigation state across tool calls. This is disabled by default — set `browser_enabled: true` to activate it.

### Browser Tools

When the browser subsystem is enabled, the agent gains these tools:

| Tool                 | Purpose                                           |
| -------------------- | ------------------------------------------------- |
| `browser_navigate`   | Navigate to a URL                                 |
| `browser_snapshot`   | Get the current page's accessibility tree         |
| `browser_act`        | Interact with page elements (click, type, select) |
| `browser_screenshot` | Capture a screenshot of the current page          |
| `browser_pdf`        | Generate a PDF of the current page                |
| `browser_tabs`       | List open browser tabs                            |
| `browser_close`      | Close a browser tab                               |

### Stateful vs Stateless Browser

<Tip>
  **Key difference:** The Computer browser subsystem maintains state (cookies, login sessions, tabs) across calls. The standalone [Browser](/docs/run/capabilities/browser) capability is stateless — each call starts fresh.
</Tip>

| Feature                  | Computer Browser (stateful)                  | Browser Capability (stateless) |
| ------------------------ | -------------------------------------------- | ------------------------------ |
| **Login persistence**    | Stays logged in across calls                 | Each call is a fresh session   |
| **Multi-step workflows** | Navigate across pages, fill multi-step forms | Single-page operations only    |
| **Tabs**                 | Multiple tabs, switch between them           | No tab management              |
| **Domain control**       | Allowlist specific domains                   | No domain restrictions         |
| **Availability**         | Agent-only                                   | Agents and MCP gateways        |

### Domain Allowlist

Use `browser_allowed_domains` to restrict which sites the browser can visit. An empty list (default) allows all domains. When set, navigation to domains not in the list is blocked.

```json theme={null}
{
  "browser_allowed_domains": ["example.com", "app.internal.com"]
}
```

## Resource Limits

| Limit                | Value            | Notes                 |
| -------------------- | ---------------- | --------------------- |
| Default SSH Timeout  | 30 seconds       | Per command execution |
| Browser Snapshot     | 500 elements max | Truncated if exceeded |
| Browser Max Tabs     | 5                | Per session           |
| Browser Max Download | 50 MB            | Per file              |

## Cost

For current pricing details, see [Pricing](https://noorle.com/pricing/).

Monitor in **Account** > **Usage** dashboard.

## Common Use Cases

### Web Application Testing

```
Screenshot app, verify buttons, test workflows
```

### Automation

```
Automate repetitive UI tasks programmatically
```

### Data Entry

```
Fill forms and navigate multi-step processes
```

### Visual Inspection

```
Verify visual appearance matches requirements
```

## Agent Loop Pattern

Typical agent workflow:

1. **Screenshot** - See current state
2. **Analyze** - LLM processes image
3. **Decide** - LLM decides next action
4. **Execute** - Perform mouse/keyboard action
5. **Repeat** - Loop until task complete

Each iteration includes LLM context (screenshot analysis), so agent sees results of actions.

## Best Practices

### Start with Screenshot

Always capture initial state before taking actions.

### Be Explicit

Use clear instructions for agent:

```
✓ "Click the 'Save' button (blue, bottom-right)"
✗ "Save the file"
```

### Handle Errors

If action doesn't work as expected:

```
"Screenshot again to verify action completed"
```

### Use Coordinates When Possible

Provide coordinates directly when known:

```
"Click at coordinates (1850, 50)"
```

### Wait for State Changes

Allow time for UI updates:

```
"Wait 2 seconds for dialog to load"
"Take screenshot to verify"
```

## Limitations

* **Desktop/Web only** - Works with rendered interfaces
* **Not for APIs** - Use HTTP Client for APIs
* **Visual interpretation** - Relies on screenshot analysis
* **Speed** - Slower than direct API calls
* **Flakiness** - UI changes can break workflows

## When NOT to Use

| Task               | Use Instead            |
| ------------------ | ---------------------- |
| API access         | HTTP Client            |
| Quick calculations | Code Runner            |
| File operations    | Files                  |
| Database access    | HTTP Client (via REST) |

## Troubleshooting

### Screenshot is blank

* Wait for page to load
* Check window is focused
* Verify viewport size is correct

### Click doesn't work

* Coordinates may be off
* Element may not be clickable
* Try right-clicking instead
* Screenshot again to verify state

### Text not entered

* Field may not be focused
* Type more slowly
* Use keyboard navigation (Tab)
* Copy-paste if typing fails

### Agent stuck in loop

* Break task into smaller steps
* Increase wait times
* Provide more explicit instructions
* Use timeout to stop execution

## Privacy & Security

<Warning>
  Computer capability has broad system access. Only use with trusted tasks.

  * Screenshots may contain sensitive data
  * Keyboard input includes all characters
  * No automatic filtering of credentials
  * Use with caution in production
</Warning>

Best practices:

* Use dedicated user accounts
* Limit to non-sensitive applications
* Monitor screen capture content
* Disable in production where possible

## API Access

```bash theme={null}
# Execute computer action
curl -X POST https://api.noorle.com/v1/agents/{agent_id}/computer \
  -H "X-API-Key: ak-{your_key}" \
  -H "Content-Type: application/json" \
  -d '{
    "action": "screenshot"
  }'

# Click action
curl -X POST https://api.noorle.com/v1/agents/{agent_id}/computer \
  -H "X-API-Key: ak-{your_key}" \
  -H "Content-Type: application/json" \
  -d '{
    "action": "click",
    "x": 100,
    "y": 100
  }'
```

## Next Steps

* [Browser Capability](/docs/run/capabilities/browser) — Stateless browser for quick page rendering
* [Web Search - Find Information](/docs/run/capabilities/web-search)
* [HTTP Client - API Access](/docs/run/capabilities/http-client)
* [Code Runner - Process Data](/docs/run/capabilities/code-runner)
* [Creating Agents](/docs/run/agents/creating-an-agent)