Repository intelligence

achiya-automation/safari-mcp

An open-source macOS Model Context Protocol (MCP) server that connects AI coding agents to your real, already-logged-in stable Safari browser.

CLOUDM0N decision

REVIEW BEFORE ADOPTION

Trust REVIEW · 60/100

Good fit if

Developers and AI agents on macOS who need to automate browser tasks using their existing web cookies, logins, and session states.

Watch out for

Restricted exclusively to macOS operating systems.

Practical intelligence

What matters before you adopt it

Decision first · detail on demand

Problem it solves

The high CPU usage, lack of authenticated session support, and risk of bot detection/Cloudflare blocks that AI coding agents encounter when running headless browsers or standard browser drivers.

Best for

Developers and AI agents on macOS who need to automate browser tasks using their existing web cookies, logins, and session states.

Main trade-offs

Restricted exclusively to macOS operating systems.

Using the optional Safari extension requires manually toggling 'Allow Unsigned Extensions' in Safari after every browser restart.

Does not support Lighthouse audits or performance trace metrics.

Why it stands out

Features 97 native tools for comprehensive web interaction, navigation, and storage management.

Highly performant with ~5ms execution per command by reusing a persistent osascript process.

Keeps all real logins, cookies, and active browser sessions intact for the agent to inherit.

Trust & CVEs

Security evidence without the noise

Trust remains a decision signal; CVEs and scanner evidence explain what is driving the risk.

Security findings

CLOUDM0N scanner findings

Critical

High

Medium

Low

View trust evidence & security findings▼

Why this score

No trust rationale was stored for this scan.

CLOUDM0N findings

MEDIUM

Dynamic code execution pattern detected. 5 sample match(es) found.

MEDIUM

The project can spawn operating-system processes; review command construction and input handling. 5 sample match(es) found.

MEDIUM

1 lifecycle install script(s) require review.

MEDIUM

Dockerfile does not end with an explicit non-root USER.

Architecture from code9 modules · 2 edges▼

Structural evidence

Modules and dependency edges extracted from repository code. This is code evidence, not README inference.

Code files

Modules

Dependency edges

Core modules

(root)

12 files

extension

4 files

scripts

6 files

xcode

1 files

Dependency flow

scripts→(root)

Detected languages

JavaScript · Swift · Shell

Detected frameworks

MCP SDK

Architecture evidence details

flowchart TD
    %% safari-mcp — high-level architecture (DRAFT, refine me)
    n0["(root) · 12 files"]
    n1["extension · 4 files"]
    n2["scripts · scripts · 6 files"]
    n3["test · tests · 31 files"]
    n4["tests · tests · 1 file"]
    n5["xcode · 1 file"]
    n2 --> n0
    n3 --> n0
    class n2 infra
    class n3,n4 test
    classDef infra fill:#b35c00,color:#ffffff,stroke:#8f4a00
    classDef test fill:#499894,color:#ffffff,stroke:#397975

Evidence, security & integrations▼

Integrations

SafariClaude CodeClaude DesktopCursorWindsurfVS CodeClineContinue

Security notes

• All communications are completely local using stdio or localhost, with no data transmitted to remote or cloud networks.

• Collects no user telemetry, performance tracking, or usage analytics.

• Requires explicit macOS system-level Privacy & Security permission approvals (Accessibility, Screen Recording, Automation) to execute native UI tools.

• The postinstall script re-signs the helper binary with a stable identifier to ensure Accessibility permissions survive future package updates.

Still unknown

• The README does not specify custom configuration properties or files for managing multi-user deployments inside larger network environments.

• No performance benchmarks are provided detailing the exact memory scaling characteristics of the HTTP daemon when running multiple parallel sessions.

Adoption guidance▼

Adopt if

+ You are on macOS and want your local AI coding agent to automate tasks within active, authenticated browser profiles.

+ You need a lightweight, low-footprint web scraper that runs in the background without stealing active window focus.

Avoid if

– You run your AI development workflows on Windows, Linux, or non-macOS environments.

– You need to run performance traces, audits, or Lighthouse analysis that require Chrome DevTools Protocol integrations.

How it works & getting started▼

How it works

1.The user installs and runs the MCP server locally using npx.

2.The developer configures their preferred MCP client (like Cursor or Claude Desktop) to connect to the server.

3.The server sends commands to Safari via AppleScript or the compiled Safari Extension without bringing the browser to the foreground.

4.The agent calls tools to inspect pages via accessibility snapshot trees and perform background actions like clicks or form entry.

5.For multi-agent setups, a shared HTTP daemon handles concurrent sessions by marking and isolating individual tabs with unique IDs.

Getting started

→ Verify your macOS machine has Node.js 20+ installed.

→ In Safari Settings, go to Advanced and enable 'Show features for web developers'.

→ In Safari's Develop menu, enable 'Allow JavaScript from Apple Events'.

→ Add 'npx safari-mcp' as a command within your target IDE's MCP configuration settings.

Agent handoff

Use with any agent

JSON API

Alternatives

Nearby repositories worth comparing before adoption.

Compare top options →

veilbrowser

acunningham-ship-it/veilbrowser

Fit

Stealth browser for AI agents — real Chrome over raw CDP, no Playwright/Puppeteer. TypeScript + MCP-native. Passes sannysoft 57/57, bypasses Cloudflare.

Trust REVIEW · 70

Compare →

stealth-chrome-devtools-mcp

DevinoSolutions/stealth-chrome-devtools-mcp

Fit

Undetectable browser automation for AI agents via MCP. Stealth Chrome DevTools with smart profile management, anti-detection arg filtering, and CDP access.

Trust REVIEW · 60

Compare →

gemini-skill

WJZ-P/gemini-skill

Fit

Looking to add image generation superpowers to your AI agents? This nifty MCP server lets any MCP-compatible agent (especially OpenClaw) tap into Gemini's drawing capabilities straight through the browser. Whether you're building automation workflows or multi-agent systems that need visual output, this project bridges the gap between AI agents and browser-based image generation without the usual complexity headaches.

Trust REVIEW · 65

Compare →