🦁 Safari MCP

The browser for your coding agent.

Your real Safari, logged in — no Chrome, no heat, no headless.

Install in VS Code · VS Code Insiders · Install in Cursor

98 tools · No Chrome/Puppeteer/Playwright needed · ~5ms per command · 60% less CPU than Chrome

Quick Start · All 98 Tools · Examples · Why Safari MCP? · Architecture · Changelog

Safari MCP Demo

❌ Without Safari MCP

Your AI agent needs to browse. So it either:

• Spins up Chromium via Playwright — with no logins, no cookies, no sessions
• Uses Chrome DevTools MCP — and melts your fan running a second browser
• Relies on headless scrapers — blocked by Cloudflare, reCAPTCHA, and bot detection

✅ With Safari MCP

Your AI drives the Safari you're already logged into — Gmail, GitHub, Ahrefs, Slack, banking.

Native WebKit. ~60% less CPU. Background operation. 98 tools. One npx command. macOS only.

📰 Featured on freeCodeCamp: How to Connect Your AI Coding Agent to a Browser on macOS · HackerNoon: Reverse-Engineering React, Shadow DOM, and CSP

🍎 Apple ships an official Safari MCP — safaridriver --mcp, first in Safari Technology Preview 247 (July 2026) and in stable Safari since 27.0. It drives an isolated automation session for debugging. safari-mcp drives the real Safari you're already logged into — your cookies and sessions, in the background, with 98 tools. See the full comparison below.

Highlights

• 98 tools — navigation, clicks, forms, screenshots, network, storage, accessibility, and more
• Zero heat — native WebKit on Apple Silicon, ~60% less CPU than Chrome
• Your real browser — keeps all logins, cookies, sessions (Gmail, GitHub, Ahrefs, etc.)
• Background operation — Safari stays in the background, no window stealing
• No browser dependencies — no Puppeteer, no Playwright, no WebDriver, no Chrome
• Persistent process — reuses a single osascript process (~5ms per command vs ~80ms)
• Framework-compatible — React, Vue, Angular, Svelte form filling via native setters

In users' own words

Not solicited testimonials — quotes lifted from the public issue tracker, each linked to the thread it came from.

"I run multiple Pi sessions/subagents against my normal Safari profile in parallel. Sharing its cookies and logins is intentional." — @maxim, on running concurrent agents against a real browser

"The server has a deliberate tab-ownership model … the code is careful about this, and for the default case that's the right safety posture." — @turner-moore, on why the guards refuse to touch your tabs

"Direct local validation from the package: Safari MCP doctor 6/6." — @jrepp, who found and fixed a queue-alignment bug in the focus helper

Quick Start

Prerequisites

• macOS (any version with Safari)
• Node.js 20+
• Safari → Settings → Advanced → Show features for web developers ✓
• Safari → Settings → Developer → Allow JavaScript from Apple Events ✓

Install (one command)

npx safari-mcp

That's it — no global install needed. Or install permanently:

npm install -g safari-mcp

Configure your MCP client

All clients run Safari MCP the same way — npx safari-mcp. Pick your editor:

claude mcp add safari -- npx safari-mcp

{
  "mcpServers": {
    "safari": {
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

{
  "mcpServers": {
    "safari": {
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

{
  "mcpServers": {
    "safari": {
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

{
  "servers": {
    "safari": {
      "type": "stdio",
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

{
  "mcpServers": {
    "safari": {
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

{
  "mcpServers": {
    "safari": {
      "command": "npx",
      "args": ["safari-mcp"]
    }
  }
}

mcpServers:
  - name: safari
    command: npx
    args:
      - safari-mcp

extensions:
  safari:
    name: safari
    type: stdio
    cmd: npx
    args:
      - safari-mcp
    enabled: true

{
  "context_servers": {
    "safari": {
      "command": {
        "path": "npx",
        "args": ["safari-mcp"]
      }
    }
  }
}

brew install achiya-automation/tap/safari-mcp

git clone https://github.com/achiya-automation/safari-mcp.git
cd safari-mcp && npm install

Usage Workflow

The recommended pattern for AI agents using Safari MCP:

1. safari_snapshot        → Get page state (accessibility tree)
2. safari_click/fill/...  → Interact with elements by ref
3. safari_snapshot        → Verify the result

Element targeting — tools accept multiple targeting strategies:

Tip: Start with safari_snapshot to get element refs, then use refs for precise targeting. This is faster and more reliable than CSS selectors.

Running several agents at once

Multiple agents or subagents driving one Safari at the same time will fight over the active tab — unless you run them against a shared HTTP daemon instead of one process per client:

SAFARI_MCP_HTTP=1 SAFARI_MCP_HTTP_PORT=9225 npx safari-mcp

Then point every client at it:

{ "mcpServers": { "safari-mcp": { "type": "http", "url": "http://127.0.0.1:9225/mcp" } } }

One daemon, many sessions — and each session gets its own tab state. The server keys activeTabIndex, the ownership flag and a unique tab marker off the MCP session id, so session A physically cannot read or steer session B's tab.

Two properties make this safe rather than merely tidy:

• Tab identity is a marker, not an index. Each session stamps a unique id into the page it opens, so ownership survives navigation and survives the user reordering or closing other tabs. An index alone would silently drift onto the wrong tab.

• It fails closed. If a session's marked tab can't be re-found, every tool refuses instead of falling back to whatever tab is in front — because that tab is usually yours:

  code

  Tab tracking lost — refusing to fall back to "current tab of window"
  (would target the user's active tab). Call safari_new_tab to reopen.
  

This also drops process count sharply: ~17 node processes for 17 concurrent sessions becomes 1.

SAFARI_PROFILE stays optional — leave it unset and sessions bind to your ordinary Safari windows, cookies and logins intact. Details in docs/http-transport-design.md.

Prefer stdio (one process per agent) over a persistent daemon? That works too — isolation then comes from the process boundary itself. One caveat: if your client multiplexes agents through mcporter, mcporter caches a single MCP client for all of them — whichever transport you pick — so the server never sees distinct sessions and per-session isolation can't engage. mcporter-lanes (a pi extension by @maxim, born out of #76) fixes this upstream: each agent session gets its own daemon dir — and therefore its own safari-mcp — with an idle timeout so processes don't pile up.

Acting on a tab you already have open

By default the server touches only tabs it opened itself. Point it at one of yours and it refuses:

Tab safety: refusing "click" — current tab (https://mail.example.com/inbox) was not
opened by this MCP session. Use safari_new_tab or safari_switch_tab to target your own tab.

That default exists because early versions did click into and close people's tabs. But "read the article I'm looking at" and "fill in the form on my screen" are real, and reopening the page loses the session state that made your tab worth using. Set SAFARI_MCP_ALLOW_USER_TABS=1 and an explicit safari_switch_tab adopts the tab instead of refusing it; from then on the session works in it like one of its own, and says so:

{ "tabIndex": 3, "safeUrl": "https://mail.example.com/inbox", "note": "(user tab, opted-in)" }

What the flag deliberately does not do:

• It unlocks adoption, not the guards. Only safari_switch_tab adopts, and only the tab you named. An ordinary click or navigate still never lands on whatever tab happens to be in front — the server acts on the tab you pointed it at, not the one you wandered to.
• safari_close_tab still refuses. Closing is the one action whose cost you cannot undo, so an adopted tab is writable, never disposable. Close it yourself.
• Adoption is session-local. Nothing is written to the shared ownership file, so it ends with the session rather than leaking to the next process on the machine.

safari_doctor prints the flag's state, and every operation on an adopted tab logs (user tab, opted-in) — so "why did it touch my tab" has an answer instead of being a mystery. Default off; set it only for agents you want working inside your own browsing session. Designed in #92.

Environment variables

SAFARI_MCP_RAISE_ON_NAVIGATE=1 is for agents whose whole point is showing you a page — a voice assistant answering "open YouTube", a demo driver. Everything else should leave it off: by default Safari MCP works in the background and hands focus back to whatever app you were using, so an agent can drive a page while you keep typing somewhere else.

Tools (98)

Security

Safari MCP runs locally on your Mac with minimal attack surface:

Safari MCP vs Alternatives

Tip: Use Safari MCP for daily browsing tasks (95% of work) and Chrome DevTools MCP only for Lighthouse/Performance audits.

vs Apple's Official Safari MCP (safaridriver)

In July 2026 Apple shipped an official Safari MCP server built on safaridriver — first in Safari Technology Preview 247, and in stable Safari since 27.0. That's great validation for the category — and it's built for a different job. Apple's server drives an isolated WebDriver automation session for debugging; safari-mcp drives the real Safari you're already logged into.

On macOS 27.0 with Safari 27.0, safaridriver --help lists --mcp and its tools/list answers with 17 entries (verified 2026-09-15); Safari 26.5.2 had no --mcp (verified 2026-07-23). Check your own machine with safaridriver --help | grep mcp.

Running both? Apple's setup guide registers its server under the name safari-mcp (claude mcp add safari-mcp -- /usr/bin/safaridriver --mcp) — the same name as this package. Give one of them a different name, e.g. claude mcp add safari-webdriver -- /usr/bin/safaridriver --mcp, so your client doesn't overwrite one with the other or leave you guessing which server answered.

About that navigator.webdriver row: the WebDriver standard requires every remote-controlled session to report true, so Apple's server does. safari-mcp evaluates JavaScript in an ordinary tab, so it reads false (checked on Safari 27.0, 2026-09-17). That is the difference if a site refuses WebDriver traffic. It is not an anti-detection feature — nothing else about the browser is disguised, and your real logins are in use, so treat every page as if it knows exactly who you are.

When Apple's server is the right pick: you specifically want a clean-room, WebDriver-standard session for compatibility debugging on Safari 27 or later. For everything else — daily automation on the browser you're already signed into, in the background — safari-mcp is built for exactly that.

Why Safari MCP and Not the Other Safari MCP Projects?

There are several "safari-mcp" projects floating around. Here's how they compare:

TL;DR — if you want the most complete Safari MCP with the smoothest install, the most tools, and active maintenance, this is the one.

Architecture

Safari MCP uses a dual-engine architecture — the Extension is preferred for speed and advanced capabilities, with AppleScript as an always-available fallback:

Claude/Cursor/AI Agent
        ↓ MCP Protocol (stdio)
   Safari MCP Server (Node.js)
        ↓                    ↓
   Extension (HTTP)     AppleScript + Swift daemon
   (~5-20ms/cmd)        (~5ms/cmd, always available)
        ↓                    ↓
   Content Script       do JavaScript in tab N
        ↓                    ↓
   Page DOM ←←←←←←←←←← Page DOM

Key design decisions:

• Dual engine with automatic fallback — Extension is preferred; if not connected, AppleScript handles everything seamlessly
• Persistent Swift helper — one long-running process instead of spawning per command (16x faster)
• Tab-indexed operations — all JS runs on a specific tab by index, never steals visual focus
• JS-first approach — typing, clicking, file upload all use JavaScript events (no System Events keyboard conflicts)
• No activate — Safari is never brought to foreground

Safari Extension (Optional)

The Safari MCP Extension is optional but recommended. Without it, ~80% of functionality works via AppleScript alone. The extension adds capabilities that AppleScript cannot provide:

What the Extension Adds

When do you need the extension? If you're automating modern SPAs with closed shadow DOM (e.g., Reddit), sites with strict Content Security Policy, or framework-heavy editors (Draft.js, ProseMirror, Slate).

Installing the Extension

The extension requires a one-time build with Xcode (free, included with macOS).

Note for npm users: The xcode/ directory is not included in the npm package. Clone the GitHub repository to build from source.

Prerequisites: Xcode (install from App Store — free)

# 1. Clone the repo (the npm package does not include the Xcode project)
git clone https://github.com/achiya-automation/safari-mcp.git
cd safari-mcp

# 2. Build a clean, Apple Development-signed app
# If needed, first choose your Team in Xcode → Signing & Capabilities.
xcodebuild -project "xcode/Safari MCP/Safari MCP.xcodeproj" \
  -scheme "Safari MCP (macOS)" -configuration Release \
  -allowProvisioningUpdates clean build

# 3. Locate the signed app
APP_PATH=$(find ~/Library/Developer/Xcode/DerivedData/Safari_MCP-*/Build/Products/Release -name "Safari MCP.app" -maxdepth 2 | head -1)

# 4. Re-sign safari-helper with the Apple Events entitlement
# (helps macOS surface the TCC Automation prompt reliably)
codesign --sign - --force --entitlements safari-helper.entitlements safari-helper

# 5. Open the app (needed once so Safari registers the extension)
open "$APP_PATH"

Do not ad-hoc re-sign Safari MCP.app with codesign --deep --sign -. That replaces the Apple Development identity on the wrapper and embedded extension; current Safari can then silently disable or mark the extension as removed. If a previous build was ad-hoc signed, run the clean build command above again.

Alternatively, open xcode/Safari MCP/Safari MCP.xcodeproj directly in Xcode, select your Apple ID under Signing & Capabilities, and click Run. A free personal Apple Developer account is sufficient for local use.

Then in Safari:

1. Safari → Settings → Advanced → enable Show features for web developers
2. Safari → Settings → Developer → Allow unsigned extensions (required each Safari restart)
3. Safari → Settings → Extensions → enable Safari MCP Bridge

The extension connects automatically to the first local bridge whose declared Safari profile matches its own. The default bridge ports are 9224, 9228, 9232, and 9236; a single-profile setup normally uses 9224.

Note: "Allow unsigned extensions" resets every time Safari restarts. You'll need to re-enable it in Safari → Settings → Developer after each restart. The extension itself stays installed.

Toolbar icon status:

• ON — connected to MCP server
• OFF — manually disabled via popup
• (no badge) — server not running, will auto-reconnect

macOS Permissions

Safari MCP needs these one-time permissions:

Granting Accessibility to safari-helper (required for safari_native_*)

The safari_native_click, safari_native_keyboard and safari_native_hover tools inject OS-level CGEvent events into Safari without stealing focus. macOS requires the underlying helper binary to be approved in Accessibility before those events can reach a non-frontmost window.

1. Open System Settings → Privacy & Security → Accessibility (on macOS 27 the pane is named Device Control and Data Access).
2. Click + (unlock with your password if needed).
3. Navigate to the helper binary and add it:
   • npm global install: $(npm root -g)/safari-mcp/safari-helper
   • npx / project install: ./node_modules/safari-mcp/safari-helper
   • From source clone: /path/to/safari-mcp/safari-helper
4. Make sure the toggle next to it is ON.

The postinstall script re-signs the helper with a stable identifier (com.achiya-automation.safari-mcp) so this permission survives future upgrades — without that step, every npm update would silently revoke approval because the binary's adhoc-signed identifier changes per build.

If safari_native_click reports success but the page doesn't react (no isTrusted: true click events fire), the helper is most likely missing this approval. The safari_* (non-native_) tools don't need it.

Granting Automation → Safari (important for IDE users)

macOS TCC grants Automation permission to the parent process that spawns the MCP server, not to safari-mcp itself. So you need to grant Automation → Safari to the app that runs Claude Code / Cursor / Windsurf — typically Visual Studio Code or Terminal.

If the permission dialog never appears automatically, run this command once from a Terminal that already has Automation permission:

osascript -e 'tell application "Safari" to get URL of current tab of window 1'

That call registers the Terminal app in the Automation database and then triggers the prompt for Safari. After you approve it, subsequent MCP calls from any child process chain will work.

Troubleshooting

FAQ

Works With

Safari MCP works with any MCP-compatible client:

Contributing

PRs welcome! See CONTRIBUTING.md for setup instructions.

The codebase is two files:

• safari.js — Safari automation layer (AppleScript + JavaScript)
• index.js — MCP server with tool definitions

Sponsors

Safari MCP is free and open source. If it saves you time or CPU cycles, consider supporting its development:

Your support funds:

• 🧪 Testing across macOS versions and Safari releases
• 🛠️ New tools and features
• 📖 Documentation and examples

Become the first sponsor!

Commercial Support

Need Safari MCP integrated into your product or agent stack? Achiya Automation offers:

• Priority bug fixes and custom tool development for your use case
• Integration consulting — wiring Safari MCP into production agent systems (Claude, Cursor, n8n, custom)
• Private deployment support — multi-user Safari MCP, non-standard macOS environments, CI/CD
• Training workshops for engineering teams adopting MCP-based automation

Built by the author of Safari MCP. Start a conversation →

What agents unlock with Safari MCP

When an AI agent drives Safari MCP, it gets things a headless browser can't:

• Real authenticated sessions — Gmail, GitHub, Ahrefs, Slack, banking dashboards are all already logged in
• Framework-aware form filling — safari_fill_and_submit calls React/Vue/Angular setters natively, no guessing whether input events fired
• Background operation — the agent works in parallel while you keep using your Mac
• One MCP call per workflow — safari_run_script batches navigation + clicks + extraction into a single roundtrip

The pattern holds across models: drive the browser the human already trusts — you inherit logins, cookies, extensions, and the user's exact environment in one step.

Community

6,000+ monthly npm downloads — developers are building AI agents on macOS with Safari MCP.

• GitHub Discussions — ask questions, share use cases
• Issues — bug reports and feature requests
• — start contributing

Ecosystem

Other macOS MCP servers that complement Safari MCP:

Using Safari MCP alongside Chrome DevTools MCP? Safari handles 95% of daily browsing (zero overhead), Chrome handles the 5% that needs Lighthouse or Chrome-specific traces.

Like it? Give it a ⭐

If Safari MCP saves you from Chrome overhead, a star helps others discover it:

Share on Twitter/X · Share on LinkedIn · Write about it

Listed On

License

MIT — use it however you want.

Built by Achiya Automation — automation & AI agents for business.