Apple shipped an official Safari MCP. I read all 17 tools - here's why I'm keeping mine.

Yesterday Apple did something I've been half-expecting since I open-sourced safari-mcp: they shipped their own Safari MCP server. It landed in Safari Technology Preview 247 (July 1, 2026), built by the WebKit team.

Any MCP-compatible agent - Claude, Cursor, whatever - can now connect straight to a Safari window and inspect the DOM, read the console, capture network requests, and take screenshots. I maintain a tool that does exactly this. So I did the obvious thing: I read the entire release, every one of the ~17 tools, and asked the only question that matters - do I need to change anything? Here's the honest answer.

What Apple actually shipped

Credit where it's due. The design is clean and the privacy story is excellent:

• It runs on safaridriver, Safari's built-in WebDriver binary. Enable "remote automation and external agents" in STP and connect with one command.
• It runs entirely locally. No network calls to Apple. Page content, screenshots, and console logs go straight to your agent - not to Apple's servers.
• The ~17 tools cover the real debugging loop: navigate_to_url, get_page_content, evaluate_javascript, page_interactions (click/type/scroll/hover), screenshot, list_network_requests, browser_console_messages, browser_dialogs, tab management, set_viewport_size, set_emulated_media.

If your job is "let an agent debug how a page renders in WebKit," this is a first-party, well-built way to do it. It's also the strongest possible validation that the category I've been building in is real.

The one line that decides everything

From Apple's own docs: "The Safari MCP server does not have access to your personal information in Safari (e.g. AutoFill or other browser activity)."

That sentence is the whole story. safaridriver drives an isolated WebDriver automation session - a clean-room window with a "controlled by automation" banner. It is not the Safari you're already using, with your tabs and your logins. And that's not a bug. For a debugging tool, an isolated session is the correct design. Reproducible, no personal-state leakage, standards-based. But it's the opposite of what I built.

safari-mcp was built for the other 95%

I didn't build safari-mcp to debug rendering. I built it so an agent could drive the browser I'm already signed into - Gmail, GitHub, Ahrefs, my bank - using native AppleScript + a Safari extension, in the background, without stealing focus, on stable Safari, on every Mac.

So here's the comparison I put in the README, straight up:

Two numbers I want to be honest about

I checked them on my own machine before writing this:

• safaridriver --mcp only exists in STP 247. The stable safaridriver that ships with Safari 26.5 has --port, --bidi, --enable, --diagnose - no --mcp. So today, using Apple's server means installing Technology Preview.
• The isolated session means no logins. Which is exactly the wall I built safari-mcp to get around.

So am I changing anything?

Yes - three things. None of them is "switch."

1. Positioning, not code. Apple didn't make safari-mcp obsolete; they clarified what it's for. Their server is the clean-room debugger. Mine is the "drive the browser you already trust" tool. Different jobs. I updated the README to say so out loud.

2. An opt-in safaridriver backend - later, not now. My architecture is already dual-engine (extension + AppleScript). Adding a third, opt-in WebDriver backend for people who specifically want a standards-based clean session is a natural extension, not a rewrite. It's gated on STP going stable, so it waits.

3. Steal the one thing they do better. WebDriver input is a rock-solid, officially-supported way to synthesize events. My native-input path leans on CGEvent, which gets fragile across macOS releases. safaridriver is a good model for a more stable input fallback. That's a genuine improvement I owe my users.

The takeaway

When a platform vendor ships an official version of your open-source tool, the reflex is panic. But "official" and "replacement" are different words. Read the whole thing before you react. Nine times out of ten it's built for a slightly different job than yours - and the honest comparison is better marketing than any launch post.

Apple built the clean-room debugger. I built the tool that drives the browser you're already logged into. Both should exist.

safari-mcp is open source (MIT) - github.com/achiya-automation/safari-mcp. It's native macOS Safari automation for AI agents: 96 tools, npx safari-mcp, no Chrome. More on what I build at achiya-automation.com.

What would make you reach for an isolated debugging session over your real, logged-in browser - or the other way around? I'd genuinely like to know.