# Safari 27 MCP: Let an agent use your browser—but can it replace ego-browser?

> Safari 27 lets AI agents click, type, and inspect webpages. Here is how to connect it, and how it differs from Chrome DevTools MCP and ego-browser.

---

LLMS index: [llms.txt](/llms.txt)

---

Safari 27's MCP server is not a chat box inside the browser. It **gives the AI assistant you already use a way to operate Safari**. Agents such as Claude Code and Codex can capture screenshots, read the DOM, inspect console messages and network requests—and also click, type, scroll, switch tabs, and run JavaScript in a page.

That supports a workflow like “open a site → search → follow an article → extract its content,” not just read-only debugging.

## How to connect {#setup}

Use **Safari 27 on a Mac**. You do not necessarily need macOS 27: WebKit also lists Safari 27 updates for macOS 15 and 26.

1. In Safari → Settings → Advanced, enable **Show features for web developers**.
2. Under Developer, enable **Allow remote automation and external agents**.
3. Add the MCP server in your terminal, choosing the command for your assistant:

```bash
# Claude Code
claude mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp

# Codex
codex mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp
```

Start a new assistant session, confirm the connection, and begin with a URL:

> Open https://example.com in Safari, check that it loads, look for console errors, and take a screenshot. Report your findings without changing any code.

Apple notes that **the first action should navigate to a webpage**. Listing tabs or creating one before the controlled window exists may fail. Once it opens, ask the agent to follow links, fill a test form, or check a narrow viewport.

## How is it different from Chrome? {#chrome}

The comparison here is with **Chrome DevTools MCP**, not Chrome's chat assistant.

| | Safari MCP | Chrome DevTools MCP |
| --- | --- | --- |
| Setup | Built-in `safaridriver --mcp` | Runs `chrome-devtools-mcp` through Node.js/npm |
| Browser actually tested | Safari/WebKit | Chrome/Blink |
| Inspection | DOM, screenshots, console, and network; performance metrics through page JavaScript | Similar page inspection, plus dedicated performance-trace recording and analysis tools |

**Already using Chrome? Safari is still a useful addition.** The main benefit is letting one agent inspect both browsers rather than guessing, from Chrome alone, why a layout breaks in Safari.

## Can it replace ego-browser? {#ego}

**There is overlap in everyday page interactions, but it is not a drop-in replacement.** ego-browser provides task spaces, access to existing signed-in contexts, and workflows for handing control between a person and an agent. Apple's documentation describes opening a new window marked as agent-controlled, without documenting equivalent task-space and handoff APIs.

In particular, do not assume it inherits your everyday Safari logins. Apple says MCP does not access personal Safari information; that is not a promise that it can take over an already signed-in Gmail or GitHub session. **Login reuse and session persistence still need testing.** The published tool list also does not establish parity with Ego's file-upload, download, and low-level control interfaces.

Safari MCP is worth considering for testing your site, reading pages, and performing ordinary interactions. Acting through your everyday accounts with smooth human handoff is a different requirement; “supports MCP” does not establish equivalence.

## Local execution does not mean all data stays local {#privacy}

The Safari MCP server runs on your machine and makes no network calls of its own. Captured pages, screenshots, and logs go to the connected agent, not Apple. **Whether they then reach a cloud model depends on your assistant's configuration.** Keep sensitive pages away from untrusted agents, and require human confirmation before submitting, sending, or purchasing.

The interesting development is not simply that Safari “has AI.” It is that **agents now have an official way to operate Safari—but operating webpages and taking over your everyday browser are different things.**

## Sources and verification scope {#sources}

This article is based on documentation, not a hands-on Safari 27 MCP connection test. Unverified login behavior is not presented as a supported feature.

- [9to5Mac: Safari 27 developer features and MCP](https://9to5mac.com/2026/09/17/webkit-blog-breaks-down-whats-new-with-safari-27-for-developers-including-mcp-support/)
- [WebKit: Features for Safari 27.0](https://webkit.org/blog/18325/webkit-features-for-safari-27-0/#safari-mcp)
- [Apple: Connecting an AI agent to Safari](https://developer.apple.com/documentation/safari-developer-tools/connecting-an-ai-agent-to-safari)
- [WebKit: Safari MCP tools and data boundaries](https://webkit.org/blog/18136/introducing-the-safari-mcp-server-for-web-developers/)
- [Chrome DevTools MCP repository](https://github.com/ChromeDevTools/chrome-devtools-mcp)