I let my AI agent use Safari

I asked Codex to open YouTube in Safari, search for a creator, visit their channel, and tell me their latest upload.

Once connected, it felt fast. Codex searched the site, opened the channel's Videos page, and checked that Latest was selected. It came back with the title, link, upload age, and runtime.

Safari has its own MCP server now, and I think that's pretty cool. It's part of Safari 27.0webkit.org (opens in a new tab) (відкриється в новій вкладці), released on September 17. Apple supplies the connection through safaridriver. Codex still runs the agent; Safari gives it tools to inspect and use a browser window. It also works with Claude Code and other MCP-compatible clients.

What happened in the test

The server returned structured page content with search fields, buttons, links, and element identifiers. Codex rejected optional cookies, entered the search, and opened the official channel. It read the newest video's details from the page and requested full URLs to get a direct link.

That worked without screenshots or custom JavaScript. There was one small failure along the way: clicking the text "Videos" matched multiple elements. Codex retried with a specific element identifier and reached the right page.

Getting connected took a few attempts. Safari reported that remote automation was disabled even though the setting was enabled. A later retry worked, but I couldn't establish what fixed it.

So "felt fast" describes the interaction once it worked. I didn't benchmark it or compare its speed with another browser tool.

Checking a Shottr release

I recorded another task: open X, find Shottr's account, read its release announcements, then check the official website to see what was available.

In the September 19 recording, Codex found that Shottr 1.9.2shottr.cc (opens in a new tab) (відкриється в новій вкладці) was available with macOS 27 compatibility improvements. The announcement on Xx.com (opens in a new tab) (відкриється в новій вкладці) also mentioned a bigger release with new features. That update was still upcoming, with no version number or release date given. The website listed 1.9.2, so Codex kept the released update and the promised one separate.

And a shout-out to Shottrshottr.cc (opens in a new tab) (відкриється в новій вкладці). It's an amazing piece of software for screenshots on the Mac.

Connect it to Codex

You'll need Safari 27 or later. In Safari's settings, open Advanced and enable "Show features for web developers" if the Developer tab isn't visible. Then open Developer and enable "Allow remote automation and external agents".

Register the server:

codex mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp

Start by asking Codex to open a specific URL in Safari. That first navigation creates the controlled window, which displays an agent-control banner. Asking to list or create tabs before that can return an error. Apple explains this in its connection guidedeveloper.apple.com (opens in a new tab) (відкриється в новій вкладці).

Safari Technology Preview uses a different driver path. WebKit's MCP introductionwebkit.org (opens in a new tab) (відкриється в новій вкладці) has the separate setup commands for it.

The next thing I want to try

The YouTube task was a small browsing test. What I want to try next is having the agent check my own site in Safari after changing the code.

Apple documents tools for reading page content, inspecting console messages and network requests, running JavaScript, taking screenshots, and changing the viewport. That gives an agent ways to investigate a broken layout or form beyond reading the source code.

Here's a prompt for that next test. I haven't run this website check yet:

Use Safari MCP to open my local website. Visit a blog post at desktop width and at 390 pixels wide. Check whether the navigation works and whether a long code block makes the page overflow horizontally. Inspect console errors and failed requests. Report what you observed and capture a screenshot if there is a visual problem.

A narrow desktop viewport can catch responsive layout problems. Testing on an actual iPhone would still be a separate check.

What stays local

WebKit says the MCP server runs locally, makes no network calls itself, and can't access Safari personal information such as AutoFill or unrelated browsing activity. The page content, screenshots, and logs it captures go to your agent. What happens to that data then depends on the agent and model. The privacy explanationwebkit.org (opens in a new tab) (відкриється в новій вкладці) is specific about that boundary. A local MCP server doesn't make the whole AI workflow offline.

A few more Safari 27 changes

The release overviewwebkit.org (opens in a new tab) (відкриється в новій вкладці) also has some useful changes for everyday web development:

• Customizable select controls let you style a native dropdown while keeping its built-in keyboard and form behavior.
• Scroll anchoring keeps the reading position steady when content loads above it. It's enabled by default.
• The rewritten ES module loader fixes evaluation-order and initialization problems, including issues around top-level await.
• Web Inspector shows color contrast in its color picker and exposes each request in a redirect chain.

These are documented release changes; my hands-on test covered MCP browsing. The release also expands spatial web support, adds WebAssembly JavaScript Promise Integration, and improves streams. WebKit counts 83 features and 844 fixes in total.

For me, the next step is that local website check. I want Codex to open the page it just changed and show me what works in Safari.