Safari’s built-in MCP server vs TabPilot

Safari 27 and Safari Technology Preview 247 include an MCP server, started with safaridriver --mcp, that WebKit calls “a Model Context Protocol server for web developers”. It works in WebDriver automation windows, which Apple says “always start from a clean slate” and can’t access browsing history, AutoFill data or other sensitive information. So: no logins. TabPilot is the opposite: it works in your own signed-in Safari, with per-site permission and your approval before risky actions. Apple’s server is the better tool for testing a website; TabPilot is for getting things done in your accounts.

Updated

What Safari’s MCP server is

In July 2026 WebKit announced that Safari 27 and Safari Technology Preview 247 include an MCP server: a way for AI agents to drive Safari through the Model Context Protocol. WebKit describes it as “a Model Context Protocol server for web developers”, and says any MCP-compatible client can connect to it.

It is part of safaridriver, the program Safari already uses for WebDriver testing. You start it in MCP mode with the --mcp flag, and the agent launches it as a local program. WebKit’s own example for Claude Code is:

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

Before it works, you turn on Safari’s developer features (Safari › Settings › Advanced › Show features for web developers) and then Safari › Settings › Developer › Allow remote automation and external agents.

What the agent gets is aimed at building websites. WebKit lists tools such as evaluate_javascript, screenshot and list_network_requests, and says the agent can get access to the DOM, network requests, screenshots and console output. It also says the server runs entirely on your Mac, makes no network calls of its own, and has no access to your personal information in Safari, such as AutoFill or other browser activity.

Automation windows: a clean slate

The important part is where it works. Safari’s WebDriver support runs in special automation windows. Apple’s documentation says these windows are isolated from normal browsing windows, user settings and preferences, and you can recognize them by their orange Smart Search field.

“Like a private browsing session, an automation session always starts from a clean slate.”

Apple adds that an automation session can’t access Safari’s browsing history, AutoFill data or other sensitive information that a normal browsing session has. While a session runs, Safari also puts a transparent “glass pane” over the window to catch stray clicks and typing; if you break through it, the session stops.

For testing, this is exactly right: every run starts the same, and nothing from your own browsing leaks into a test. For errands it means the agent starts signed out of everything. It can’t see your inbox, your orders or your company’s dashboard, and you can’t sign in for it halfway through without ending the session.

What TabPilot does instead

TabPilot is a Mac app with a Safari extension. The agent works in the Safari you already use, with your tabs and your logins. Because that’s far more powerful, TabPilot puts you in charge of it:

  • Site Access per website: Ask First (the default), Read Only, Read & Act or Off. A new site asks you before the agent sees anything.
  • Approval before risky actions. Buying, paying, sending, deleting and submitting wait for Allow Once or Don’t Allow in the page.
  • Secrets stay yours. Password, card and one-time-code fields are hidden from the agent when it reads a page, and it can’t type into them.
  • Prompt-injection shield. Hidden text is removed and instruction-like lines are marked as untrusted page content.
  • Take Over. For a sign-in or a 2FA code the agent hands the tab to you, and continues when you press Done.
  • You can see and stop it. A blue cursor and caption show each step, a pill says “TabPilot is controlling this tab”, and Option-Shift-P stops the agent.

TabPilot has no tool that runs JavaScript written by the agent. In a test window that’s a useful power; in a browser signed in to your bank, it’s a risk we’d rather not hand out.

The connection is different too. Safari’s server is a program the agent starts. TabPilot’s app runs a small MCP server on 127.0.0.1 that the agent reaches over HTTP with a token. Setup is one command for Claude Code: see how to connect Claude Code to Safari.

Side by side

Safari’s MCP serverTabPilot
Where it worksWebDriver automation windows, isolated from normal browsingYour normal Safari windows and tabs
Your logins, cookies and open tabsNo. Each session starts from a clean slateYes. The agent works signed in as you
Browsing history and AutoFillNo accessNot read. Password, card and one-time-code fields are hidden from the agent
Who it’s for“Web developers”, in WebKit’s wordsAnyone who wants an agent to do errands in their own browser
How the agent connectsStarts it as a local program: safaridriver --mcpTalks to the TabPilot app over HTTP on 127.0.0.1, with a token
What to turn onSafari’s developer features and “Allow remote automation and external agents”The TabPilot extension in Safari › Settings › Extensions
Per-site permissionNone described in its docsSite Access per site: Ask First, Read Only, Read & Act, Off
Approval before buying, sending, deletingNone described in its docsYes: Allow Once or Don’t Allow, in the page
You can step in mid-taskNo: interacting stops the sessionYes: Take Over for sign-ins and 2FA, then Done
Runs agent-written JavaScriptYes: evaluate_javascript is one of its toolsNo. A fixed set of tools, none runs the agent’s code
Best forTesting and debugging your own websiteEveryday tasks on sites you use

Which one to use

Use Safari’s MCP server when you’re building or debugging a website and want an agent to load pages, read the console, inspect network requests and take screenshots in a clean, repeatable window. It’s free, built in and made for exactly that.

Use TabPilot when the task lives in your accounts: reading what’s in your open tabs, filling in a form you’re signed in to, collecting order details, or doing a chore across several sites. That needs your real Safari, and it needs you to stay in control, which is what Site Access, approvals and Take Over are for.

Many developers will want both. They don’t conflict: they are separate MCP servers with different names.

Questions

Can I use both?

Yes. They are separate MCP servers with different names, so an agent can have both: Apple’s for testing a site you build, tabpilot for your everyday Safari.

Does Safari’s MCP server see my passwords?

No. Apple says automation windows can’t access browsing history, AutoFill data or other sensitive information, and WebKit says the server has no access to your personal information in Safari.

Why can’t Safari’s MCP server just use my logins?

Because automation windows are deliberately isolated. Apple describes them as starting from a clean slate, like a private browsing session, so tests aren’t affected by earlier sessions. That isolation is the point for testing, and also why it can’t do errands in your accounts.

Is TabPilot made by Apple?

No. TabPilot is an independent app with a Safari extension, sold on the Mac App Store. It is not affiliated with Apple.

Sources

More guides

TabPilot — MCP for Safari

Let your agent use your real Safari, with per-site access and your approval before anything risky.

Coming soon to the Mac App Store

One-time purchase · macOS 14 Sonoma or later · Try the demo store