# Let the Claude Code you already pay for clean up your C drive: Purrge hit 280 stars in 17 days, author has only 1 followers · GitHub open-source project

> Source: https://stringers.io/en/github/gh_RabbittuX__purrge.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> People who use AI coding tools are the most likely to fill up their disks, and "have Codex clean my C drive" is popular in the Chinese-speaking internet right now. It needs no new API key, costs nothing, and makes "AI reads only, a human checks the boxes" its selling point. With Windows support added within a week, it was most likely passed around twice by a few WeChat groups or Official Accounts.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

RabbittuX/purrge · RabbittuX

# Let the Claude Code you already pay for clean up your C drive: Purrge hit 280 stars in 17 days, author has only 1 followers

**What it is** A Chinese-language disk cleaner for Mac and Windows. When you're low on space, it reminds you. Then the Claude Code or Codex you're already logged into on your computer explains, item by item, whether each thing is safe to delete. You tick what goes, and only then does it delete.

**Why it's blowing up** People who use AI coding tools are the ones whose disks fill up fastest. "Have Codex clean my C drive" is a trend in the Chinese dev community right now. Purrge needs no new API key and costs nothing, and it makes "AI can only read, a human checks the boxes" its selling point. It added Windows within a week, and it has likely been shared twice in WeChat groups and Official Accounts.

**B** **How you could ride it** Don't build another cleaner; the top spots in that space are already taken. Take the approach instead: borrow the user's logged-in Claude Code / Codex as the engine, and build only the permission boundaries, the interface, and the background presence. Apply it to scenarios nobody has covered yet, like clearing Docker / node_modules, organizing the Downloads folder, or cleaning the agents' own session caches, and publish real test results on RedNote and WeChat groups.

Grade B: the approach is clear and copyable, and few people know about it yet. But the disk-cleanup space already has Mole at 70k stars and Khazix's skill at 20k stars. Purrge itself has only 200–300 stars, the sources of those stars can't be traced, and big-company AI assistants are building this into system features. The star ceiling for a follower is low. Hardly anyone knows about it yet

Data (updated daily) · Repo created 2026-10-20 · 295 stars · 7 days +132

GitHub · stars

Data 2026-10-26 → 2026-11-08 (14 days, 14 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

An unknown Chinese developer turned the popular prompt "have Codex help me clean up my disk" into a small tool you can install, keep running in the background, and that lets the AI look but not delete. It never made trending, and almost nobody reposted it publicly. Even so, it climbed to about 280 stars in 17 days. Disk cleanup is already crowded. What's worth studying is the approach: the product ships no model of its own and uses the agent the user has already logged into on their machine.

### An issue from a D drive user

On the morning of Nov 1, someone opened issue #5 on Purrge's GitHub titled "Support monitoring and scanning multiple disk partitions (Windows multi-drive users are only covered on C)". He runs Windows 10 22H2, with the system on C and his data and projects on D. He went through the 0.3.4 source code: space monitoring only checks `%SystemDrive%`. Of the 22 Windows cleanup rules, all start from `%LOCALAPPDATA%` or `%APPDATA%`. The large-file scan is hardcoded to the 6 folders Downloads, Desktop, and Documents. As a fallback he suggested: "If multi-volume monitoring is too costly short term, exposing point #2 (user-configurable scan paths) would solve most of the pain. Thanks!"

The repo was only created on Oct 20, and the Windows version arrived on Oct 24. An outsider had already read the source this closely, and that same day was the day it gained the most stars: +45.

### What it is

Purrge (purr is a cat's purr, purge means to clear out) is a disk cleaner for macOS and Windows, with a README and interface in Chinese. It does three things:

- **Watches remaining space in the background** : it checks on a schedule and only pops up a card asking "Want me to take a look?" when space falls below a threshold. This step doesn't call any AI and uses no quota.

- **AI judges each item** : it scans for caches, installers, large files, and old projects, then hands them to Claude Code, Codex, Tencent WorkBuddy, or DeepSeek, whichever you're already logged into locally. The AI explains what each item is, how long it has been idle, and what happens if you delete it.

- **You tick, it deletes** : caches that can be rebuilt are checked by default, and your own files are unchecked by default. The AI has no delete permission, and anything deleted goes to the Trash by default.

The command-line version is open source (non-commercial license) and can be installed via Homebrew, a one-line script, or PowerShell Gallery. On Nov 5 it also released a closed-source Mac GUI version with a menu bar and a floating desktop ball, downloadable only from the official site and free of charge.

### Why it could take off

**The demand is created by AI coding tools themselves.** Through 2026, two kinds of content kept circulating in the Chinese-speaking world. One was "Codex is quietly eating your hard drive" and "two sessions ate 115GB" (RedNote). The other was "have Codex clean up C drive". Tutorials for it appeared on Zhihu, CSDN, and YouTube, and 36 Qi also ran a review that opened with someone on X using Codex to clean a MacBook, "saving up to 500GB". People who use Claude Code and Codex every day are exactly the ones whose disks fill up most often and who are most used to letting AI do the work.

**No new API key, and no charge.** It calls the Claude Code or Codex you're already logged into and uses the quota from your own subscription. Similar tools like Pinkbin require an API setup, and CleanMyMac requires a subscription. For someone who already pays for Claude Max or ChatGPT Plus, Purrge has zero barrier to entry. It also supports WorkBuddy and DeepSeek, which serves Chinese users without an overseas subscription.

**It turned "afraid AI will delete the wrong things" into a selling point.** Under RedNote posts about AI disk cleaning, someone commented, "I'm rather conservative, I don't dare use AI for fully automatic cleanup." Purrge's answer is very specific: with Claude, all tools are turned off; Codex runs in a read-only sandbox; only paths, sizes, and idle times are sent to the AI, never file contents; and files modified after the scan are skipped automatically. It makes clear exactly who presses the delete key.

**Proactive reminders that a prompt can't provide.** Skills and prompts only run when you remember to run them. Purrge waits in the background and only comes looking for you when space runs low.

**Fast fixes.** The Windows version shipped within 4 days of the repo's creation, and from Oct 29 to Nov 5 it released 5 fix versions back to back. The error reports in the issues are very specific: a GBK decoding crash on Chinese Windows, a WorkBuddy port already in use, and a newer Claude Code version where `--output-format json` started returning an array. Replies typically came the same day, with fixes shipped the same day, and users who submitted PRs later appeared as commit authors.

### Where the stars came from: mostly untraceable

| Date | stars | What happened that day |
|---|---|---|
| 10-20 | Repo created | Back-to-back v0.2.0–0.2.2, Homebrew tap created the same day |
| 10-24 |  | v0.3.0 adds Windows, WorkBuddy, and DeepSeek |
| 10-26 | 30 |  |
| 10-27 | 60 | +30 in one day |
| 10-28 | 93 | +33 in one day |
| 11-01 | 163 | +45 in one day, the day's peak; the multi-partition issue |
| 11-02 | 207 | +44 in one day |
| 11-05 | 262 | Mac GUI version 1.0 |
| 11-06 | 275 |  |

Figures come from the daily GitHub records (11-06). With 28 forks, the forks-to-stars ratio is about 10%, which is normal. The 12 issues and PRs came from at least 7 accounts, and all describe errors from real environments. They don't look fabricated.

The author RabbittuX's account was only registered in August, has 1 followers, no bio, and no linked X account. Searches on X, V2EX, linux.do, WeChat Official Account search, and Douyin turn up no posts mentioning Purrge. RedNote has just one post from Oct 29, with 0 likes and 1 saves. GitHub aggregator sites have indexed it, but only after it had already taken off.

So the sources of the stars can only be inferred. The two jumps of thirty to forty stars a day around 10-27/28 and 11-01/02 look like a single share by some Official Account or WeChat group. Almost everyone who filed issues was a Chinese Windows user, matching the audience of "C drive full, find an AI". Another part likely came from GitHub's internal search, where it is tagged claude-code, codex, macos, and disk-cleanup. This is inference, not confirmed.

### Who it's competing with

The space is very crowded, and the biggest players were there first:

| Competitors | Scale | How Purrge differs |
|---|---|---|
| tw93/Mole | 74k stars, +739 in the last 7 days | Rule-driven deep cleaning and uninstalling, no AI |
| Khazix's storage-analyzer skill | In a repo with 22k stars | Has to be invoked manually inside an agent, with no installer, background alerts, or interface |
| Pinkbin | 1,412 stars | The closest approach: drag a folder to the AI and ask whether it can be deleted. Only a Windows installer, runs on an API, no proactive reminders |
| PureMac, MacSai, burrow, MangoDisk, and others | 1.5k–7.6k stars | "Open-source CleanMyMac", rule-based |
| Tencent Lemon, CleanMyMac, DaisyDisk | Long-established | One-click cleaning, doesn't explain each item |
| Tencent Marvis | big tech | A system-level AI assistant. One sentence in chat can sort and scan for installers and caches, with a graphical interface (a review reposted by 36 Qi) |
| A pile of skills and small tools (Moppet, Pika Clean, CleanMyPc, etc.) | From 0 to a couple hundred stars | Most do only one thing: AI judgment, or one-click cleanup |

There's also a commercial app with the same name, purrge.app (one-time purchase $14.5–29, no AI). It has no connection to this project, but it competes for search results.

Each of these pieces has been built by someone before. Purrge is one of the few Chinese open-source tools that puts "use the subscription you already have + AI that only reads + a human checks the boxes + proactive reminders" into one installable tool. That's why it got shared and kept, but a count of two to three hundred stars also shows it hasn't broken out yet. Compared with Mole's 70k stars, it's a niche tool with good word of mouth.

### If it were you

**Verdict** : building another AI disk cleaner will struggle to beat Purrge's two to three hundred stars, and the top spots are already split between Mole and Khazix. What's worth taking is the approach: "no model of your own; borrow the Claude Code / Codex the user is already logged into as the engine; build only the permission boundaries, the interface, and the background presence." Then run it again in a scenario nobody has covered yet.

**Who it suits** : developers who use Claude Code / Codex every day, can write a desktop or command-line tool, and know the Chinese developer community.

**Who it doesn't suit** : people who want to make money from it (this kind of open-source project earns stars and reputation, not revenue), and people with no distribution channel who won't post about it themselves. Most of Purrge's stars have untraceable origins, so don't expect GitHub's internal search to send people your way automatically.

**How to do it:**

- Change the scenario, not the skin. The same structure, "AI judges read-only, a human checks the boxes, the tool executes", can move to clearing Docker images and node_modules, organizing Downloads and screenshots, or cleaning the caches and sessions of Claude Code / Codex themselves (the crowd behind "two sessions ate 115GB" on RedNote).

- Make safety the selling point. Put what the AI can and can't see, who presses the delete key, and how to undo it, line by line, on the first screen of the README.

- Make installation one line: Homebrew, PowerShell Gallery, and a one-line script, and add Windows in the first week. Most of Purrge's issues came from Windows users.

- Do what it hasn't done well: issue #5 asked for multi-partition support and custom scan paths, which it still lacked in early November.

**How to get users** : add dual-theme screenshots to the Chinese README and tag it claude-code and codex. On release day, post a real test on RedNote, Jike, and WeChat groups titled something like "I cleared xx GB with Codex", with a screenshot of the review page. Purrge's two jump-shaped spikes show that a single share in a community can bring thirty to forty stars.

**When to move** : the sooner the better, ideally before the end of 2026. Big-company AI assistants like Tencent Marvis can already clean up disks through chat. Once system-level assistants become common, "have AI take a look" will become a built-in OS feature.

**Signs you've misread it** : if Purrge's daily star growth drops to single digits and levels off around 300 stars, the trend has been picked clean by the leaders. If Claude Code or Codex ships its own "clean cache / free up space" command, or tightens how third-party programs can use login sessions, this borrow-the-subscription approach needs to be reassessed.

[GitHub repo ↗](https://github.com/RabbittuX/purrge)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| tw93/Mole | Open source | 2025-09 | 74,015 stars, +739 in the last 7 days | Rule-driven deep cleaning and uninstalling, without AI judgment. Several hundred times Purrge's size. |
| khazix-skills / storage-analyzer (Khazix, digital life) | Open source | 2026-04 | In a repo with 21,995 stars, and a RedNote post calling it "the most popular computer cleaning skill online" | A Skill you invoke manually in an agent, with no installer, background alerts, or GUI |
| cccyd2003-qwq/pinkbin | Open source | 2026-05 | 1,412 stars | Also hands a folder to AI for judgment, but ships only a Windows installer, runs on an API, doesn't use the Claude/Codex subscriptions users already have, and has no proactive alerts |
| PureMac / MacSai / burrow / purge-app / MangoDisk | Open source | 2026-04 to 08 | 1.5k–7.6k stars | An open-source take on CleanMyMac, cleaning by rules. burrow gives Agents MCP access, the opposite direction. |
| CleanMyMac / Tencent Lemon Cleaner / DaisyDisk | Big-company features | Years | Commercial mainstream; Lemon 6.4k stars | A one-click black box: CleanMyMac is subscription-based and doesn't explain each item |
| purrge.app (a commercial product with the same name) | Other | 2026-10 | 2 votes on Product Hunt, daily rank #160 | Same name, different company. $14.5–29 one-time purchase, no AI. It will compete with Purrge for search results. |
| Just asking Codex / Claude Code to clean the disk (prompts, various small skills) | Big-company features | Since the first half of 2026 | Plenty of tutorials on Zhihu, CSDN, YouTube, and 36 Krypton | You have to write prompts yourself every time, and you're afraid the AI will delete the wrong things. Purrge productized it: AI read-only, manual checkboxes, proactive alerts. |

Where users come from

- Inference **GitHub in-site discovery (topics: claude-code/codex/macos/disk-cleanup)** : no external posts found on X, V2EX, linux.do, WeChat Official Accounts, or Douyin, yet stars keep climbing. The repo is under 60 days old and hasn't made trending.

- Inference **Sharing through Chinese private channels (WeChat groups / Official Accounts / Jike)** : two step-ups of +30 to +45 around 10-27 and 11-01. Issue and fork accounts are almost all Chinese users, mostly on Windows. No public source found in search.

- **Confirmed** **Xiaohongshu** : the account 老白AI拆解局 posted a #Purrge note on 10-29, with 0 likes and 1 saves. Not enough reach to matter.

- **Confirmed** **Package managers and website (conversion)** : Homebrew tap (10-20), PowerShell Gallery (10-24), website dmg (11-05). The README and landing page are both polished.

- **Confirmed** **GitHub aggregator / SEO sites** : ossradar, aisignal (11-02), vibecrowd, skillsllm, and hlist auto-indexed it, but all after stars were already rising.

## RabbittuX/purrge: why it got about 280 stars in 17 days

> Research date 2026-11-06. Data from our github database (repo_profile / repo_day, daily data from 10-26 onward), GitHub API (releases, commits, issues), Jina / Exa / Brave search, TikHub Xiaohongshu / Douyin search, x-radar and the x database. Stargazer details weren't retrieved this time (GitHub account rate-limited, and that endpoint can't be called unauthenticated), so it's not possible to verify where each star came from.

### One line

Purrge is a Chinese-language macOS / Windows disk cleaner that bundles no model of its own: **it has the user's already-logged-in Claude Code / Codex / WorkBuddy / DeepSeek judge, read-only and item by item, "what is this and what happens if I delete it"** . The user then checks items, and the tool does the deleting. It also watches remaining space in the background and alerts only when it drops below a threshold. It works because of the very hot wave in the Chinese scene of "use Codex / Claude Code to clean my C drive or Mac" through the 2026 year: it turned that hand-rolled prompt setup into a product with **safety boundaries, an install path, and proactive alerts** . Almost no spread shows up on the public web, so the stars most likely came from Chinese channels search engines don't index (WeChat groups / Official Accounts / Moments) plus GitHub in-site discovery. I couldn't pin down which one.

### 1. Data

| Date | stars | fork | What happened that day |
|---|---|---|---|
| 10-20 | Repo created |  | v0.2.0–0.2.2 released the same day; homebrew-tap created the same day |
| 10-24 |  |  | v0.3.0 added Windows, WorkBuddy, DeepSeek; v0.3.1 published to PowerShell Gallery |
| 10-26 | 30 | 6 |  |
| 10-27 | 60 (+30) | 8 |  |
| 10-28 | 93 (+33) | 14 |  |
| 10-29 | 97 | 16 | The 1th external issue (Claude Code connection test crashes on Windows) |
| 10-31 | 118 | 18 | v0.3.3 / 0.3.4 fix Windows issues |
| 11-01 | 163 (+45, peak) | 21 | Issues #5 and #6: Chinese Windows GBK crash, and a user requesting multi-partition support |
| 11-02 | 207 (+44) | 23 |  |
| 11-05 | 262 | 28 | Released Purrge for Mac 1.0 (closed-source GUI, download only from the website) and v0.3.5 |
| 11-06 | 275 (API real-time 281) | 28 |  |

Source: github repo_day; release times from the GitHub API. gain_7d=170, max_day_gain=45(11-01).

**Is the star count real?** : The fork-to-star ratio is about 10%, which is within the normal range. 12 issues/PRs come from at least 7 different accounts (jiangwanyutao, lih288504-design, oopsdoes, Ashy-bot-hub, Su-1003, NaIWaniubi, and others), and the content is real errors from real environments, such as `to_trash` failing to decode GBK on Chinese Windows, WorkBuddy hitting EADDRINUSE on its port, and the newer Claude Code `--output-format json` output coming back as an array. NaIWaniubi submitted a PR and later showed up as a commit author; Su-1003 filed 4 reliability issues in one go, with fixes attached. These are signs that real people are installing and using it, not the pattern of inflated numbers. Also, most of the issues are about Windows, which suggests the Windows version 10-24 shipped brought in a sizable share of users.

**Author** : GitHub account RabbittuX, registered 2026-08-04, with only 1 followers, no bio, and no linked X account. The website is hosted at static.yellowrabbit.online. The author has no personal influence to lean on.

### 2. Competitors

This space is very crowded, and the biggest players came earlier:

| Competitors | Form | Launched | Scale | How Purrge differs |
|---|---|---|---|---|
| tw93/Mole | Open-source CLI + native app | 2025-09 | 74,015 stars, +739 in the last 7 days | Rule-driven deep cleaning and uninstalling, no AI judgment. Its name recognition and scale are hundreds of times Purrge's. |
| storage-analyzer in KKKKhazix/khazix-skills | Agent Skill | 2026-04 | The whole repo has 21,995 stars (the author, 数字生命卡兹克, is a top Chinese AI blogger). Xiaohongshu has a note calling it "the hottest computer cleaning skill online" (88 likes, 9-27). | Users have to call it themselves from inside an Agent. No installer, background monitoring, or GUI. Purrge turned this whole approach into a product. |
| cccyd2003-qwq/pinkbin | Open-source Tauri desktop app | 2026-05 | 1,412 stars | The closest idea (scan the disk, drag folders to an AI to ask whether they can be deleted, send to trash), but it only ships a Windows installer and calls an API (the README lists an API relay sponsor). It doesn't use the user's existing Claude / Codex subscription and doesn't alert proactively. |
| momenbasel/PureMac, iliyami/MacSai, caezium/burrow, jithin-sabu/purge-app, harry0703/MangoDisk | Open-source native Mac cleaners | 2026-04 to 08 | 1.5k–7.6k stars | The "open-source CleanMyMac" approach, relying on rules rather than AI. burrow gives Agents MCP access, the opposite direction: an Agent calls it, not the other way around. |
| Tencent/lemon-cleaner (Tencent Lemon), CleanMyMac, DaisyDisk | Big tech / commercial | Years | 6.4k stars / major commercial players | One-click cleaning is a black box. CleanMyMac requires a subscription. |
| purrge.app (a different commercial product with the same name) | Paid Mac app, $14.5–29 one-time purchase | Blog since 2026-10-15 | 2 votes on Product Hunt, daily rank #160 | Same name, different company. Unrelated to RabbittuX and doesn't use AI. It will compete with Purrge for search results. |
| A lot of skills and small tools: Jet3-ai/mac-storage-cleanup (3 stars), hisunfei/broom (0 stars), CSZHK/CleanMyPc "Atlas" (210 stars), Moppet on V2EX (10-26), MacAutoClean, Pika Clean on Juejin (10-26), Moon Clean / 轻灵 on Xiaohongshu | Skills / desktop pets / apps | Appeared over the past 2026 year | Most have only a few to about two hundred stars | Most are either "a single skill" or "rules plus one-click cleaning." Few do AI judgment, proactive alerts, allowlists, undo, and multiple engines all together. |
| The "just have Codex / Claude Code clean the disk" approach | Big-company features / prompts | Since the first half of 2026 | Covered by Zhihu, CSDN, YouTube, and 36 Krypton ("AI cleared 500GB") | Nothing to install, but you write the prompt yourself every time and worry the AI will delete the wrong things. Purrge addresses that fear with the read-only boundary. |

Sources: github repo_profile (11-06); GitHub API; Xiaohongshu via TikHub search; Exa / Jina search results (V2EX, Juejin, Zhihu, and 36 Krypton links at the end).

### 3. Growth tactics: what was found and what wasn't

- **Owned channels** : the README is in Chinese and well made (screenshots in both themes, a comparison table, FAQ). The landing page is static.yellowrabbit.online/purrge. Install options cover the Homebrew tap, a one-line script, and PowerShell Gallery. This is strong conversion work, not acquisition.

- **Chinese community posts** : no posts mentioning Purrge found on V2EX, linux.do, Juejin, Zhihu, or Sogou WeChat search. Neither x-radar nor our x database has a single mention of purrge. Xiaohongshu only turned up one post, "Worried about deleting the wrong things when clearing AI cache? Try this tool #Purrge" (老白AI拆解局, 10-29,0 likes, 1 saves), which won't drive meaningful traffic. Nothing found on Douyin.

- **GitHub aggregator and SEO sites** : ossradar, aisignal (11-02 wrote auto-generated analysis, and the performance numbers in it are made up), vibecrowd, gitofanton, skillsllm, hlist.cc, forkfree all index it, but they all wrote about it only after it started gaining stars. They're results, not causes.

- **In-site GitHub discovery** : topics tagged claude-code, codex, ai, macos, disk-cleanup, and claude-code / codex are terms Chinese developers search every day in 2026. The repo was under 60 days old and never hit trending, so the homepage exposure didn't come from trending.

- **Inference (unconfirmed)** : the three steps 10-27, 11-01, 11-02 jumping +30 to +45 look like it was forwarded once by a WeChat Official Account or a community. Windows users in China make up most of the users, matching the crowd searching "C drive full, find an AI for it". The author had no public accounts before, so it was probably the author or people around them posting in channels search engines can't crawl, like WeChat or Jike.

### 4. Why it took off

1. **Demand amplified by the AI coding tools themselves** : in 2026 the Chinese internet was full of "Codex / Claude Code filled up my disk" (Xiaohongshu: "Codex is quietly eating your hard drive", "two sessions ate 115GB") and "have Codex clean my C drive" (Zhihu, CSDN, YouTube, 36 Kr). The people using AI coding tools are also the ones whose disks fill up most and who are most used to letting AI do the work, and Purrge lands right on that group.

2. **No new API key, no fee** : it calls the Claude Code / Codex you're already logged into, and the quota comes from your own subscription. Pinkbin-style tools need an API key, CleanMyMac needs a subscription; for people who already pay for Claude Max / ChatGPT Plus, Purrge has zero barrier. It also hooks into Tencent WorkBuddy and DeepSeek, catering to domestic users without overseas subscriptions.

3. **Turns "afraid AI will delete the wrong things" into a selling point** : the AI can only read (Claude turns off all tools, Codex uses a read-only sandbox), sends only paths, sizes, and idle times, never file contents; user files are never pre-checked by default, and deletions go to the Trash by default; files changed after the scan are automatically skipped. This directly answers comments on Xiaohongshu like "I'm rather cautious and don't dare let AI run fully automatically".

4. **Proactive alerts** : a background job periodically checks remaining disk space, calls no AI and spends no quota, and only pops up "Want me to take a look?" when it drops below a threshold. Skills and prompts can't do this; they all rely on the user remembering to run them.

5. **Fast iteration, Windows added within a week** : 10-20 repo created, 10-24 Windows support and published to PowerShell Gallery, 10-29 to 11-05 shipped 5 fix releases in a row, 11-05 Mac GUI released. External issues were typically answered and fixed the same day, and users who submitted PRs were turned into contributors.

### 5. Differentiation, and whether it's really why it grew

The differentiation is a combination of three things: **using the Agent subscription you already have, AI read-only plus manual checkbox selection, and proactive alerts when space runs low** . Each one alone has been done by someone: Mole and PureMac have reversible cleanup, khazix's skill has AI judgment, Pinkbin has drag-to-AI questions. But packaging them into a small product you can install and keep running is among the most complete in open-source Chinese projects. This combination **is why it got shared and kept** , but its scale (two to three hundred stars) shows it hasn't broken out: compared with differentiation, reach looks more like the ceiling. Next to Mole's 70k stars and khazix's 20k stars, it's currently just a niche tool with good word of mouth.

### 6. What to borrow

- "Piggybacking on the Agent subscription users already have" is a common playbook for 2026 AI micro-tools: the product itself pays no model fees and charges nothing, and only handles orchestration, permission boundaries, and UI. Any "let AI judge for me, then execute" scenario works this way, such as organizing photos, cleaning up browsers, clearing Docker images, or clearing node_modules.

- The safety story needs to be concrete: spell out what the AI gets and doesn't get, who presses the delete key, and how to undo it.

- Ship the CLI first, expand to Windows and package managers within a week, and release a GUI in two weeks; keep the GUI closed-source and open-source the CLI under a non-commercial license, leaving room to charge later.

### 7. Visibility and copycats

- Visibility: **few people know about it yet** . No discussion can be found on X, V2EX, linux.do, WeChat Official Accounts (via Sogou), or Douyin, and Xiaohongshu has only 1 posts with 0 likes; only the GitHub aggregator sites auto-index it.

- Copycats: no product directly copied it; hwl513782273/purrge and sz6660/purrge are just mirror forks with 0 stars. There are many works in the same niche from the same period (Moppet, Pika Clean, both 10-26), but nothing shows they're copying it. There's also a commercial product with the same name, purrge.app, which is unrelated.

### 8. Uncertain

- Who shared the two steps 10-27 and 11-01 couldn't be found; the stargazer timeline and account profiles weren't retrieved.

- How many people are actually using it is unknown: download counts for the Mac 1.0 dmg and install counts for Homebrew and PowerShell Gallery weren't retrieved (the PSGallery page came back with no numbers).

- The author's identity is unclear, and who is behind yellowrabbit.online is also unknown.

- The license says PolyForm Noncommercial, the Mac App is closed-source, and whether it will charge later is unknown.

### Sources

- [github.com/RabbittuX/purrge](https://github.com/RabbittuX/purrge) (README, releases, issues #1–#11)

- [static.yellowrabbit.online/purrge/index.html](https://static.yellowrabbit.online/purrge/index.html)

- [www.aisignal.dev/analysis/rabbittux-purrge](https://www.aisignal.dev/analysis/rabbittux-purrge) (11-02)

- [ossradar.net/repo/RabbittuX/purrge](https://ossradar.net/repo/RabbittuX/purrge), [vibecrowd.ai/repos/rabbittux--purrge](https://vibecrowd.ai/repos/rabbittux--purrge)

- Xiaohongshu note 6ae3089f000000002700bc7d (Lao Bai AI breakdown room, 10-29); "the hottest computer cleanup skill on the whole internet" (Ai Daily by Ah Mu, the shovel-officer, 9-27); "Codex is quietly eating your hard drive" (Yuan Xiao'er learns AI)

- [m.36kr.com/p/3831719105619842](https://m.36kr.com/p/3831719105619842) (AI cleared 500GB)

- [www.v2ex.com/t/1252028](https://www.v2ex.com/t/1252028) (Moppet, 10-26), [juejin.cn/post/7700823297483030563](https://juejin.cn/post/7700823297483030563) (Pika Clean, 10-26)

- [github.com/cccyd2003-qwq/pinkbin](https://github.com/cccyd2003-qwq/pinkbin), [github.com/tw93/Mole](https://github.com/tw93/Mole), [github.com/KKKKhazix/khazix-skills](https://github.com/KKKKhazix/khazix-skills)

- [purrge.app/](https://purrge.app/), [hunted.space/product/purrge-make-room-keep-what-matters](https://hunted.space/product/purrge-make-room-keep-what-matters)

- Our GitHub repo_profile / repo_day (2026-11-06)

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.