# Test in the built-in browser

Preview a build beside your session, let the agent drive the page, pick elements, open DevTools and capture proof in the desktop app.

> **Desktop app only**
>
> The built-in browser is part of the [desktop app](/setup/desktop-app). On the web, a session's **Browser** button opens the session in the desktop app, or the download page if you don't have it yet.

The browser opens as a pane beside your session's chat. You and the agent look at the same page: you test by hand, the agent drives it with its own tools, and what you find goes into the chat or the run as proof.

For a guided run of the whole flow, see [test a staging link in the built-in browser](/walkthroughs/browser-qa).

## Open a page

In a session, choose **Browser** in the header.

Choose **Open a page** and pick where to go:

- **Your local app**: the app in this session's folder. **Run** starts it, and **Open** opens it once it is _Running_.
- **Storybook**: the Storybooks found in the folder.
- **Preview or staging link**: paste a link your team shared, then **Open link**.
- **Recent pages**, or **Any address**, typed in the address bar.

The address bar searches DuckDuckGo when you type words instead of an address. Each tab has **Back**, **Forward** and **Reload**. **Wide view** gives the page more room than the chat. What the session made, such as a prototype or a document, opens here as a tab too.

The browser keeps a few things away from pages. A page can't open EasyCode's own addresses, and when a page tries to open more tabs, EasyCode blocks them and tells you. A dialog from a page says it comes from the page, not from EasyCode.

## Test on another device size

Open **Tools** on the toolbar. Under **View**:

- **Device size** picks a device, **Fit the panel**, or a **Custom size**. **Rotate** turns it, and **Touch** makes it a touch screen.
- **Appearance** shows the page in **Light** or **Dark**.
- **Page tools** opens DevTools, docked under the page.

Under **Tab**, **Activity** lists what the agent did on the page, step by step.

## Let the agent use the page

Your assistant gets browser tools in the session. It reads the page with [`browser_snapshot`](/reference/mcp-tools#browser-snapshot), opens addresses with [`browser_navigate`](/reference/mcp-tools#browser-navigate), acts with [`browser_click`](/reference/mcp-tools#browser-click) and [`browser_type`](/reference/mcp-tools#browser-type), and takes masked screenshots with [`browser_screenshot`](/reference/mcp-tools#browser-screenshot). The [MCP tools reference](/reference/mcp-tools) lists the rest.

The agent asks before it uses a site: _Let Claude Code use northwind.dev?_ Answer **Allow** or **Don't allow**. When a page may be signed in as you, it asks separately, _Let Claude Code use this page with your logins?_ Answer **Allow this time**, **Always allow here** or **Don't allow**. **Always allow here** shows only when the choice can be remembered.

While the agent holds a tab, a bar says so:

- **Show** brings that tab forward.
- When you use the page yourself, the agent pauses. **Let the agent continue** hands it back.
- **Stop** stops the agent using the browser until you let it again.

The agent can't type into password, one-time code or card fields. It asks you to fill them in.

## Point at what's wrong

- **Pick an element** marks parts of the page. Click a part, or drag over an area, and write a note on each: _What should change here?_ **Send _N_ notes** adds them to your message. You can add up to 20 at a time.
- **Send to agent** adds the page's errors and failed requests to your message.

When the project has source stamps turned on, a picked element also carries its file name. Your assistant can set this up with [`browser_stamp_setup`](/reference/mcp-tools#browser-stamp-setup).

## Capture proof

From **Tools**, under **Capture**:

- **Screenshot** takes the page, and **Screenshot an area** takes the part you drag over.
- **Record this page** records the tab. **Show clicks and shortcuts** draws a ring on each click and names keys like Enter. Typed text is never shown, and private details on the page are covered while it records.

The **Proof** panel collects your recordings and the agent's checks, which it runs with [`proof_run`](/reference/mcp-tools#proof-run). Proof is attached to the session's run. When part of a page couldn't be checked for private details, the proof is _held on this computer_ until you look it over and choose **Attach anyway**. In a private session, proof stays on this computer and nothing is uploaded. **Copy PR block** copies the proof for a pull request's description.

A recording you made has **Report a bug**: steps in EasyCode's words, your comments, and **Copy report**, **Send to agent** or **Create bug** on the task you are testing.

If your org has the design system, **X-ray** in **Tools** marks the system components and one-offs on the page. See [X-ray and conformance](/design-system/x-ray-and-conformance).

## Keep sign-ins apart

Each workspace has its own **Browser profile**, under **Tools** › **Tab**. Sign-ins stay in that workspace and aren't shared with anyone. The profile lists the sites you always allowed, and selecting one removes it. **Clear this profile** signs you out of every site in this workspace and forgets what its pages saved. Other workspaces are not touched.

> **When the browser is off**
>
> EasyCode staff can switch the browser off for an org with [`browser`](/reference/kill-switches-and-modules#browser). The desktop app then stops opening it for that org. See [kill switches and modules](/reference/kill-switches-and-modules).
