A Developer Had Claude Install 900 Apps and Built a Free Mac Cleaner
Paid Mac cleaners charge for heavy interfaces and hundreds of extra features. One developer asked Claude to install 900+ popular apps one by one on a clean macOS virtual machine, a "computer inside a computer" with no junk, and to record which files remain after each app runs. Only findings confirmed through repeated checks went into the database. The outcome is MacWashing, a free open-source cleaner that works without an internet connection.
AI-processed from Habr AI; edited by Hamidun News
Mac cleaners cost money, look bloated, and try to do everything at once: they clean the system, act as an antivirus, and, as the author jokes, they hardly play the accordion. One developer on Habr decided there was no reason to pay for this. He asked the Claude AI model to install more than 900 popular apps, one at a time, on a clean virtual machine running macOS (a "computer inside a computer" with no junk to begin with), and to record which files remained after each program ran. From these observations he built a rule database for MacWashing, a free open-source cleaner that works without an internet connection.
How Claude built the junk database
The central problem for any cleaner is knowing what can be deleted. Guessing from folder names in Library is risky, because one wrongly deleted folder can break a program or wipe someone's settings. The author took a different approach and observed how the apps actually behaved.
Imagine you want to find out what mess each tenant leaves in an apartment. The simplest method is to move them in one at a time into a completely empty flat and see what appears. A clean virtual machine plays exactly that role. Everything that appears after installing and using a program can be tied to that specific app: caches, logs, temporary files, and leftovers after uninstalling. The virtual machines ran several macOS versions, because paths and system behavior differ from one version to the next.
Next, according to the author, came the most important stage: every result was checked, then checked again, and again. Only what was confirmed as junk went into the database. For a program that deletes files, this is the only sensible approach: a missed gigabyte is annoying, but deleting an important file can lead to data loss.
Why it works without the internet
The description indicates the utility makes no network requests. It is reasonable to assume the AI worked at the database-building stage, not during cleaning: the finished, verified rules live on the user's computer. For the user, this brings three benefits: no delays, no list of installed programs sent anywhere, and no fees for model queries on every launch.
For developers, this is a useful pattern. The expensive and not always predictable work of the model can be moved into an offline pipeline, which produces a fixed artifact (a rule database) that can be distributed without any AI involved. The tokens are spent once rather than on every user. The author does not disclose how much time and money went into building the database, so the economics of the approach cannot yet be assessed in numbers.
What remains open
The author writes that the results significantly outperform what is available on the market. But the text contains no comparative measurements: how much space is freed, how many apps each competitor recognizes, or whether there were any false deletions. The only known details are that testing covered several macOS versions and classic sets of programs for five roles: designer, programmer, engineer, photographer, and video editor.
There is also an obvious limitation. 900 apps is a substantial set of popular software. Niche programs and new versions released after the database was built will be covered less well, and the database will need to be updated the same way. The author himself says he made the program for his own use and did not plan to publish it.
What to do with this
If you use a Mac, take a look at the in4in-dev/macwashing repository on GitHub. The code is open, so you can read and verify the deletion rules before trusting them with your disk. Before the first run, it is sensible to make a Time Machine backup.
If you build your own tools, take the approach as a note: observe behavior in an isolated environment, turn the results into verifiable rules, and double-check them several times. That way the model helps gather the data, and you are not left making risky decisions on the fly.
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
- Nov 21, 2026Qwen3-0.6B Fine-Tuned on a Mac mini: Errors Dropped to 3 out of 30 from 9
- Nov 21, 2026The Programmer Stopped Being an AI Agent Dispatcher: Multica Takes Over the Handoff Routine
- Nov 21, 2026Developer Skips Pinecone: Searching 72,450 Code Chunks in 50 ms
Key Questions Answered
How Claude built the junk database?
The central problem for any cleaner is knowing what can be deleted. Guessing from folder names in Library is risky, because one wrongly deleted folder can break a program or wipe someone's settings. The author took a different approach and observed how the apps actually behaved. Imagine you want to find out what mess each tenant leaves in an apartment. The simplest method is to move them in one at a time into a completely empty flat and see what appears. A clean virtual machine plays exactly that role.
Why it works without the internet?
The description indicates the utility makes no network requests. It is reasonable to assume the AI worked at the database-building stage, not during cleaning: the finished, verified rules live on the user's computer. For the user, this brings three benefits: no delays, no list of installed programs sent anywhere, and no fees for model queries on every launch. For developers, this is a useful pattern.
What remains open?
The author writes that the results significantly outperform what is available on the market. But the text contains no comparative measurements: how much space is freed, how many apps each competitor recognizes, or whether there were any false deletions. The only known details are that testing covered several macOS versions and classic sets of programs for five roles: designer, programmer, engineer, photographer, and video editor. There is also an obvious limitation.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.