Safari 27.0 ships an MCP server, so a coding agent can drive the browser itself

Checking how a page renders in Safari has meant a loop of screenshots, a terminal and a lot of describing. Safari 27.0, documented by WebKit on September 17, 2026, removes that loop: the release ships an MCP server that gives Claude Code, Codex or another agent control of a Safari window. The agent reads the DOM, watches network requests, takes its own screenshots and reads console output, which means it can check its work instead of asking you what went wrong.

What the MCP server gives an agent, and what it does not

The server exposes a real Safari window rather than a headless approximation. WebKit lists verifying user state in forms and checkout flows, comparing computed styles and layout against other browsers, testing for accessibility problems such as missing labels and poor contrast, and reading navigation timing and resource load times. It is switched on in Settings, under the Developer pane, with the option that allows remote automation and external agents.

The privacy boundary is stated plainly: the server runs entirely on your machine, makes no network calls of its own, has no access to your personal information in Safari, and any captured data goes to the agent you are running rather than to Apple. That matters, because a browser automation tool is by construction pointed at whatever you happen to be signed in to.

The select element finally takes styling

Applying appearance: base-select switches the real <select> to the new control, with user-agent styles that WebKit redesigned over the summer rather than leaving developers to rebuild the basics. Those defaults cover the drop shadow, 4px rounded corners, touch-friendly line height, hover states, chevron and checkmark, light and dark modes, forced colors and disabled states. WebKit took the design to the CSS Working Group so that every browser can converge on the same starting point.

New pseudo-elements ::picker-icon and ::checkmark target parts of the control that could not be styled before, HTML elements can be placed inside each <option>, and <selectedcontent> controls what is shown for the currently selected option. Because it is still a real form control, keyboard navigation, screen reader support, form submission, validation and change events keep working without any JavaScript.

Layout, images and the quality pass

Scroll anchoring is the change most readers will feel without knowing its name. When something loads in above the viewport, Safari now adjusts the scroll position so the content stays where it was, instead of pushing the reader down the page. It needs no code on your side; overflow-anchor defaults to auto, and none opts out where a page genuinely needs the old behavior.

• The stretch keyword works with width, height, the min and max variants and flex-basis, and unlike width: 100% it accounts for margins, so it does not overflow.
• sizes="auto" on a lazily loaded image lets the browser compute the size from the actual layout width, so you no longer have to predict it.
• The HTML <model> element, on visionOS for a year, now works in Safari on iOS, iPadOS and macOS, with multiple <source> files and an <img> fallback.
• CSS gains alpha() for adjusting only the alpha channel, color-mix() with more than two colors, image(<color>) as a real image layer, light-dark() accepting images, the :heading pseudo-class and the revert-rule keyword.