Claude-to-Safari-MCP

Claude-to-Safari-MCP

A small collection of Model Context Protocol servers for Claude.

MCP	What it does	Where
safari-mcp	Drive your real macOS Safari (open URLs, read pages, run JS, screenshot). Zero dependencies, single file.	safari-mcp.js
google-docs-mcp	Read + edit Google Docs and Sheets, including charts and structural ops (T3). Multi-account aware.	google-docs-mcp/

Each MCP has its own setup instructions — see the linked file/folder.

---

claude-safari-mcp

Give Claude hands and eyes inside Safari. Zero dependencies, one file.

claude-safari-mcp is a Model Context Protocol server that lets Claude (Desktop, Code, or any MCP client) drive Safari on macOS — open URLs, read page text, run JavaScript, manage tabs, capture screenshots. The whole server is a single self-contained Node.js file. No npm install. No build step.

Why

The web is still where most work happens. Browser-use frameworks like Playwright and Puppeteer launch their own headless Chromium, which means:

• You log in again. And again.
• Your sessions, cookies, and 2FA prompts live somewhere your agent can't see.
• It costs hundreds of MB of RAM per run.

This server piggybacks on the Safari you already have open, with the cookies and sessions you already trust. Claude works in your browser, not a sandboxed copy of it.

Tools exposed

Tool	What it does
open_url	Navigate the current tab to a URL
new_tab	Open a URL in a new tab
get_current_tab	Read URL + title of the active tab
list_tabs	List every tab across every window
activate_tab	Switch to a specific tab
close_current_tab	Close the active tab
reload_page	Reload the active tab
go_back / go_forward	History navigation
get_page_text	document.body.innerText of the active tab
get_page_html	Full outer HTML of the active tab
run_javascript	Execute arbitrary JS in the active tab and return the result
search	One-shot web search (google / duckduckgo / bing / brave)
take_screenshot	PNG of the front Safari window to a path you choose

Install

curl -O https://raw.githubusercontent.com/<your-username>/claude-safari-mcp/main/safari-mcp.js
chmod +x safari-mcp.js

That's it. Node 18+ is the only requirement — no npm install.

One-time macOS setup

1. Enable JS-from-AppleEvents (only if you want run_javascript): Safari ▸ Settings ▸ Advanced ▸ "Show Develop menu in menu bar" then Develop ▸ "Allow JavaScript from Apple Events".
2. Automation permission: the first call will trigger a system prompt asking your terminal (or the Claude app) to control Safari. Approve it. You can revisit at System Settings ▸ Privacy & Security ▸ Automation.

Wire it up to Claude

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "safari": {
      "command": "node",
      "args": ["/absolute/path/to/safari-mcp.js"]
    }
  }
}

Restart Claude Desktop. You'll see "safari" in the tools menu.

Claude Code

claude mcp add safari -- node /absolute/path/to/safari-mcp.js

Any other MCP client

Point it at node safari-mcp.js over stdio. Protocol version 2024-11-05.

Try it

In Claude:

Open hacker news in Safari and tell me the top three story titles.

Take a screenshot of the current tab and save it to ~/Desktop/page.png.

In the active tab, run document.querySelectorAll("a").length and tell me how many links are on the page.

How it works

Everything routes through Apple's osascript — AppleScript for tab/window control, JXA (JavaScript for Automation) for richer reads like list_tabs. JSON-RPC 2.0 over stdio is implemented by hand, which is why there are no dependencies.

The whole protocol surface fits in ~30 lines at the bottom of safari-mcp.js: initialize, tools/list, tools/call, plus the notifications/initialized ack. If you want to add a tool, drop a new entry in the TOOLS array and you're done.

Limitations

• macOS only (Safari + osascript). For cross-platform browser control look at puppeteer-mcp instead.
• run_javascript is gated behind the Develop-menu setting above. Apple requires explicit opt-in.
• The server controls your real Safari. Be thoughtful about what you let an agent click.

Roadmap ideas

• Reader-mode text extraction (cleaner than innerText)
• Bookmark + Reading List read APIs
• Per-tab screenshot (not just front window)
• Form-fill helpers built on top of run_javascript
• Resource subscriptions: stream URL changes back to Claude

PRs welcome. The whole codebase is one file easy to grok, easy to fork.

License

MIT.