Desktop Sandboxes
Real browser sandboxes that run your app end-to-end for visual validation and QA.
Desktop Sandboxes give the agent an isolated virtual desktop with a real browser. The agent can navigate your app, interact with UI elements, take screenshots, and run E2E tests — all within a sandboxed environment that streams live to your dashboard.
How It Works
Each Grind mode session runs inside an isolated sandbox with:
- A full Linux desktop environment
- A real browser (Chromium) for web interaction
- Git, build tools, and shell access
- Your repository pre-cloned and ready
The sandbox desktop streams to the Desktop tab in the session panel via VNC, so you can watch the agent work in real time.
Viewing the Desktop
When a Grind session is running, click the Desktop tab in the sandbox panel to see the agent's live desktop.
| State | What You See |
|---|---|
| Active session | Live desktop stream with interactive controls |
| Connecting | Loading indicator while the VNC connection establishes |
| Session complete | Desktop remains viewable for 1 hour after the session stops |
| Recording | Video playback of the full session for review |
Taking Control
You can interact with the sandbox directly:
- Click Take control to enter interactive mode (fullscreen)
- Use your mouse and keyboard to navigate, click, and type in the sandbox
- Click Hand back control or press Esc to return to view-only mode
This is useful for debugging, manually testing something the agent built, or guiding the agent through a tricky UI flow.
Browser Tools
When browser support is enabled, the agent has access to these tools:
| Tool | Description |
|---|---|
| Navigate | Go to a URL in the browser |
| Click | Click an element on the page |
| Type | Type text into the focused element |
| Fill | Fill a form input field |
| Scroll | Scroll the page viewport |
| Get State | Read the current page DOM and take a screenshot |
| Close | Close the browser tab |
The agent uses these tools to interact with web applications visually — verifying UI behavior, testing user flows, and validating that code changes work correctly in a real browser.
Use Cases
- E2E validation — Agent builds a feature, then opens the app in the browser to verify it works
- UI bug fixes — Agent reproduces a visual bug, fixes it, and confirms the fix in the browser
- Form testing — Agent fills out forms, submits them, and checks responses
- Screenshot comparison — Agent captures before/after screenshots for visual regression
- Full-stack QA — Agent starts a dev server, runs the app, and tests end-to-end
Enabling Browser Support
Browser support is automatically available in Grind mode sessions. For CLI usage, add the --browser flag:
paragon --lra "test the login flow end-to-end" --browserNext Steps
Grind Mode
Learn about the planning and execution engine that powers sandbox sessions.
Sessions
Create and manage agent sessions.