Apple Ships Safari MCP Server to Let AI Agents Debug Sites in the Browser

Apple has added a Model Context Protocol (MCP) server to Safari Technology Preview 247, letting AI coding agents connect directly to a live Safari window. Rather than a developer manually reproducing a bug, screenshotting it, and describing it in a prompt, the agent can now inspect the actual rendered page — reading the DOM, computed styles, network requests, console output, and screenshots — and iterate on fixes without the usual window-hopping. Any MCP-compatible client works; setup is a one-line claude mcp add or codex mcp add command (or an mcp.json entry) pointing at safaridriver with a --mcp flag, after enabling remote automation in Safari’s developer settings.

Beyond basic debugging, Apple pitches the server for cross-browser compatibility testing, performance analysis via navigation and resource timing, accessibility checks (missing labels, bad ARIA, poor contrast), and verifying UI state across flows like checkout. The agent is expected to invoke the tools autonomously from simple prompts such as “find bugs on my site in Safari,” without being told to use the server explicitly.

On privacy, Apple states the server runs entirely locally, makes no network calls of its own, and has no access to AutoFill or other browser data. Captured page content, screenshots, and logs flow to the connected agent rather than to Apple — meaning downstream handling depends on whichever model the developer trusts. That trust caveat is the real story: this hands third-party AI agents deep visibility into whatever site is loaded, so the security posture rests entirely on the chosen client, not on Safari itself.