Visual Testing

Headless test suites tell an agent the build is green. They don’t show it that the button is off-screen, the modal is white-on-white, or the chart never rendered. Visual testing gives the DevBox a real graphical desktop, so the agent can open your running app, see the rendered UI, and click through it the way a person would.

What the Display capability ships

Turn on Display and the DevBox gets a real graphical desktop:

• A real graphical desktop — a headed screen with a window manager, streamed live to your browser so you can watch.
• VS Code — a full IDE running on the desktop, so the agent (and you) can open files, run the editor, and use its tooling.
• A real Chromium — an actual browser, not a headless shell, with the Claude extension already wired in so the agent can drive the page.

This isn’t a screenshot service bolted on the side — it’s a real screen inside the DevBox that the agent shares with you.

Two ways an agent tests visually

Engine	How it sees the screen
Claude	Computer-use — Claude drives the real desktop directly: it looks at the screen, moves the pointer, clicks, and types, reacting to what’s actually rendered.
Codex	Playwright — Codex runs Playwright against the live Chromium render, scripting the browser and asserting on the real DOM/page.

Either way the agent is acting on a real render, not a headless approximation — so it catches the visual and end-to-end bugs a headless run sails right past.

Enabling Display

1. When launching a DevBox (Quick Launch, a Run, or a Schedule), pick an image that includes the Display capability — the combined image ships it.

2. Launch. The DevBox comes up with the graphical desktop running alongside the usual terminal.

3. Open the Display view from the container to watch the desktop stream live in your browser.

Note

Display adds a graphical stack (X server, window manager, browser) to the DevBox, so it wants a little more memory than a pure-terminal box. Pick a tier with headroom — see Cloud workers.

Watching it live

The desktop streams to your browser. You can watch the agent open your app and click through it in real time, and you can take over — grab the pointer and keyboard yourself, poke at the same screen the agent is using, then hand control back. It’s a shared screen, not a recording.

On Windows

Visual testing works on Windows DevBoxes too. A real Windows VM with the desktop means the agent can verify your app actually renders on Windows — useful when “it works on Linux” isn’t the same as “it ships.”

Inside Runs and Schedules

Visual testing isn’t a separate mode — it’s a capability of the DevBox, so it works anywhere a DevBox does:

• In an interactive Run, you can watch the agent open the app and step in when something looks off.
• In a Schedule or autonomous Run, the agent uses the desktop unattended — e.g. an overnight job that builds a feature and visually verifies the UI before opening a PR.
• In a multi-agent run, one agent in the relay can be the one that opens the app and checks the render.

Related

• Multi-agent runs — a relay where one agent tests visually
• Windows Containers
• Runtime Catalog — the visual desktop in the DevBox
• Runs · Schedules