Apple shipped an MCP server inside Safari Technology Preview 247 on July 1, and it’s one of the more practical things to land for web developers this week. Connect any MCP-compatible agent — Claude, Cursor, GitHub Copilot, Gemini CLI — to a live Safari window, and the agent can inspect your DOM, capture console errors, analyze network requests, run JavaScript, take screenshots, and check accessibility. No prompting required. No manual copy-paste of DevTools output. The Safari MCP server lets the agent see what Safari sees.
The tool targets a specific, tedious class of bug: works everywhere except Safari. Chasing those issues has always meant toggling between DevTools, reproducing the problem, describing it to an agent, and iterating. The Safari MCP server collapses that loop — agents observe and act directly.
What the Safari MCP Server Gives Your Agent #
The server ships with 17 tools. The highlights: browser_console_messages
to pull buffered logs, list_network_requests
and get_network_request
for full request details (headers, body, timing), evaluate_javascript
to run code in page context, screenshot
for PNG captures, and page_interactions
for DOM actions — click, type, scroll, hover, keyPress. There’s also set_emulated_media
for testing responsive breakpoints and dark mode without touching system settings.
Apple’s suggested prompts are deliberately simple: “Find bugs on my site in Safari.” “How accessible is my site in Safari?” That’s the point — hand the agent a URL and a task, and it handles the observation. 9to5Mac’s coverage of the release highlighted how this removes the manual back-and-forth between browser inspection and AI prompting.
Security-wise, the server runs entirely on your local machine and makes no external network calls. It does not expose AutoFill data or browser history — only tab content your agent actively requests. What happens to that data afterward depends on which agent and model you’re using, which Apple sensibly notes in the docs.
How to Connect the Safari MCP Server #
Install Safari Technology Preview 247 (a separate download from stable Safari). Enable developer features via Settings > Advanced, then turn on remote automation under Settings > Developer. Then register the server with your MCP client.
For Claude Code, that’s one command:
claude mcp add safari-mcp-stp -- "/Applications/Safari Technology Preview.app/Contents/MacOS/safaridriver" --mcp
For any other MCP client, add this to your mcp.json or equivalent config:
{
"safari-mcp-stp": {
"command": "/Applications/Safari Technology Preview.app/Contents/MacOS/safaridriver",
"args": ["--mcp"]
}
}
That’s it. The server is built on top of safaridriver — Safari’s existing WebDriver binary — with a new --mcp
flag. No additional installs, no separate process to manage, and it works with any MCP-compatible client: Cursor, Copilot, Codex, Gemini CLI.
Related:[WebMCP: Make Your Website AI-Agent-Ready in Chrome 149]
Safari MCP vs Playwright MCP vs Chrome DevTools: Pick the Right Tool #
The live-session advantage is real. Unlike Playwright or headless Chrome, the Safari MCP server connects to your actual running Safari — logins, cookies, and session state all intact. Testing authenticated flows without setting up test credentials is genuinely convenient. Federico Viticci (MacStories) called it “the most ergonomic browser MCP for agents to date” and switched back to Safari from Chrome specifically because of it.
There’s also an Apple Silicon argument. Community testing shows WebKit using roughly 60% less CPU than Chromium-based automation tools. If you’ve watched your MacBook fans spin up during Chrome DevTools MCP sessions, Safari MCP is noticeably quieter.
However, the caveats matter. This is Safari Technology Preview, not stable Safari — there is no announced timeline for a stable release. It is macOS only; WebKit doesn’t run on Linux or Windows. And if you need Lighthouse audits, performance traces, or memory heap snapshots, Chrome DevTools MCP still has the richer toolset. According to a detailed comparison of browser MCP servers, most developers who’ve tested all three end up keeping them all: Safari for live debugging, Playwright for cross-browser CI, Chrome DevTools for performance work.
What Apple Being Last Actually Signals #
Apple is the final major browser vendor to ship an official first-party MCP server. Microsoft built Playwright MCP. Google built Chrome DevTools MCP. Anthropic built Puppeteer MCP. MCP itself was donated to the Linux Foundation’s Agentic AI Foundation in December 2025 and now sees 97 million monthly SDK downloads — up from 2 million at launch in 16 months. Apple’s arrival completes the picture: every major browser now has an official agent interface built on the same protocol.
That’s worth noting not as a milestone for its own sake, but because it changes expectations. Browser tooling for AI agents is no longer experimental or third-party. It’s a first-class surface that every vendor is now competing on. Safari joining means developers can reasonably expect agent-native debugging across all major browsers going forward.
Key Takeaways #
- Apple’s Safari MCP server (shipped in Safari Technology Preview 247, July 1) gives AI agents direct access to live Safari sessions: DOM inspection, console logs, network requests, JavaScript execution, screenshots, and accessibility audits.
- Setup requires Safari TP 247, two settings toggles, and one CLI command. Any MCP-compatible client works — Claude, Cursor, Copilot, Gemini CLI, Codex.
- The live-session model means agents can test authenticated flows without test credentials, and WebKit uses roughly 60% less CPU than Chrome-based tools on Apple Silicon.
- It’s complementary to Playwright MCP and Chrome DevTools MCP, not a replacement — use it for Safari-specific debugging, not cross-browser CI pipelines.
- Limitation: Safari Technology Preview only, macOS only, no stable release timeline announced.