macOS System Optimizer: AI That Knows What’s Safe to Delete

The Problem

Mac users accumulate gigabytes of hidden junk — Xcode derived data, browser caches, orphaned app support folders, duplicate files, forgotten startup items. Existing cleanup tools either take a brute-force approach (delete everything) or require users to manually evaluate hundreds of items.

The core issues:

• Fear of breaking things — users don't know which files are safe to remove and which are critical system files
• Hidden storage hogs — ~/Library folders, Xcode caches, and Docker images silently consume 20-50GB
• No intelligent guidance — existing tools show file lists with no context about what each item does or whether it's needed
• Developer-specific bloat — Xcode derived data, CocoaPods caches, simulator images, and old archives need specialized handling

What We Built

A native macOS system optimization app built entirely in SwiftUI with MVVM architecture. The key differentiator: an integrated LLM recommendation engine that analyzes scanned files and tells the user exactly what each item is, whether it's safe to remove, and why.

8 cleanup modules

• Startup Items — identify and disable unnecessary launch agents and daemons slowing down boot time
• Unused Apps — detect applications that haven't been opened in months, with size breakdown
• Hidden Folders — surface ~/Library junk, app support orphans, and forgotten caches
• Large Files — find storage hogs buried deep in the filesystem
• Browser Caches — clear Safari, Chrome, and Firefox caches with per-browser breakdown
• Xcode Cleanup — derived data, archives, simulators, and device support files (often 20-40GB)
• Duplicate Files — hash-based detection of exact duplicates across the filesystem
• System Junk — logs, crash reports, tmp files, and other system detritus

AI-powered recommendation engine

This is what makes the app different from every other cleanup tool. After scanning, the LLM analyzes each found item and returns a structured recommendation:

• Safe to remove — caches, temporary files, old logs. No risk.
• Review recommended — items that might be needed (e.g., an app support folder for a rarely-used app). The AI explains what it is and lets the user decide.
• Essential — keep — system-critical files that should never be touched. The AI explains why.

The LLM prompt builder constructs context-aware prompts that include the file path, size, last modified date, and surrounding directory structure. The response parser handles flexible JSON output with fallback strategies for robustness.

Dual LLM backend

• Claude API — cloud-based, highest accuracy, best for nuanced file classification
• Ollama (local) — fully offline, no data leaves the machine, good for privacy-sensitive users

Users can switch between backends based on their preference for accuracy vs. privacy.

Security Scanner

Beyond cleanup, the app includes a security scan module that checks for common macOS security issues: unsigned applications, suspicious launch agents, open network ports, and misconfigured permissions.

Technical Approach

The app follows strict MVVM architecture. Each of the 8 modules has its own ViewModel that handles scanning logic, progress reporting, and state management. The LLM integration sits as a separate service layer with a protocol-based design — ClaudeLLMService and OllamaLLMService both conform to the same interface, making backend switching seamless.

The LLM prompt builder constructs structured prompts that give the model enough context to make accurate recommendations. The response parser handles the inherent variability of LLM outputs through flexible JSON parsing with multiple fallback strategies — critical for a production app where a parsing failure would mean no recommendation at all.

Results

• 8 specialized modules covering every major source of macOS storage waste
• AI recommendations that eliminate guesswork — users know exactly what each file is and whether it's safe
• Dual LLM support — cloud accuracy when needed, full offline privacy when preferred
• 100% native SwiftUI — fast, lightweight, no Electron overhead
• Security scanning built-in, not bolted on — catches issues other cleanup tools ignore

Key Takeaway

System cleanup tools have existed for decades, but they all share the same flaw: they show you a list of files and expect you to know what's safe to delete. That's like giving someone an X-ray and expecting them to read it.

By integrating LLM-powered analysis directly into the scanning pipeline, this app turns a stressful decision (“will deleting this break something?”) into a confident action. The AI reads the X-ray for you.