Everyone is teaching AI to reverse-engineer. He built a replica tool that's "look only, never take apart": where CRAB's 200 stars came from
What it is An open-source MCP toolkit: one AI only observes the software from the outside and writes a spec, and a separate, isolated AI rewrites it from only the spec and tests it against the original. The original software's code is never touched.
Why it took off It rides the REA "AI reverse-engineering replica" wave and directly addresses the criticism that "one agent doing everything isn't a clean room." There was no external catalyst; the stars came from MCP directories, GitHub topics, and the author linking 4 of his repos to each other the same week.
CHow you can follow Don't build another replica toolkit. Find the point the next hot wave gets criticized most for and build the opposite. Put honest test results in the README, ship several small libraries in a week, and cover PyPI, Homebrew, MCP directories and awesome lists. If you stay on the replica track, build a lightweight "behavior-only" skill.
- Data (updated daily) · Repo created 2026-10-22 · 234 stars · 7 days +115
GitHub · stars Data 2026-10-27 → 2026-11-08 (13 days, 13 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
In Oct 24, Austin Huang, a developer in Sydney, pushed the first commit of a new GitHub repo. That month's hottest topic on GitHub was "let AI reverse-engineer and replicate closed-source software," led by REA, which already had over 400k stars. His project goes the other way: the first line of the README reads "Rebuild it from how it behaves, never from its code." It looks only at how the software behaves and never touches a line of its code. Two weeks later it had about 210 stars. The number is small, but where it came from is clear, and one person can learn to copy it.
A wave, and the biggest fight inside it
Some background first. morluto/rea (Reverse Engineer Anything) is an MCP tool that lets coding agents like Claude Code call Ghidra, IDA and JADX to take apart binaries, Electron apps and APKs, then rebuild their functionality. It had about 37k stars in Oct 9, rose to 180k in Oct 13, and by Nov 6 had 436k stars and 108k forks (GitHub, 11-06). Right after, a batch of "clean-room replica" projects appeared: WordCraft replicating Word (17k stars), FigCraft replicating Figma, replicas of Auth0 and Salesforce, and a 8KB lightweight alternative, leanrea (4,019 stars).
The wave was big and the arguments were loud. The most cited criticism is in the "Legal and controversy" section of the awesome-reverse-engineer-anything list: one agent both analyzes and rebuilds, which "doesn't meet the personnel separation traditional clean-room development requires." One blog put it more bluntly: to make an agent follow clean-room rules, you have to isolate it at the filesystem and network layers, not rely on prompt instructions. REA's own SECURITY.md also admits it is not a sandbox.
CRAB turned these criticisms into features, one by one.
What it is
CRAB stands for Clean Rebuild of Anything's Behavior. It's written in Python under the Apache-2.0 license and installs with pip install crab-cleanroom. Its workflow is split into two sides:
- The observer side: an observer agent treats the target software (a command-line tool, web app, or Android app) as a black box. It records what the software accepts and rejects, its default values, how it handles edge cases, and how pages navigate. It deliberately triggers server 500, network drops, and bad data to see how the software copes. What it collects is only behavior: exit codes, output, accessibility trees, interface shapes. Code, decompiled output, and stack traces are never collected.
- A gate: observations are written as rules that can be verified. Before anything is passed through, the gate checks for copied original text and real names (real names get swapped for codenames). In strict mode, each release requires a human to sign off in the terminal.
- The rewrite side: a second agent sees only this spec and writes its own implementation. It can run tests against the finished product for comparison with the original, but it only gets back pass/fail and never sees anything from the original.
Both sides can be locked in containers. Isolation is enforced by the container, and every crossing is logged to a signed, hash-chained audit log.
The README publishes results on a hidden test set that the model and the observer never saw: one command-line tool 397/397, the three most recent web apps 129/130, 126/130, 125/127, and an Android app 113/118. Failures are published too: the open-source app Flipt only got 19/55. It also openly admits that when tested on well-known tools like jq, part of the result is the model's memory, and it would have done just as well without the spec. When every feed is full of "I cloned Photoshop with AI," writing it this way is what makes engineers trust it.
How the stars grew
| Date | What happened | star |
|---|---|---|
| 10-22 | Repo created | — |
| 10-24 | First commit, roughly 50–60 stars over two days (inferred by working backward from daily counts) | — |
| 10-27 | 51 | |
| 10-28 | Released to PyPI as teardown-cleanroom 0.3.0 | 69 |
| 11-01 | 119 | |
| 11-03 | Renamed to CRAB 0.4.0 | 153 |
| 11-05 | Single-day peak +20 | 186 |
| 11-06 | fork 34–35 | 196 (live about 210) |
(Daily GitHub star counts, logged from 10-27 onward; earlier numbers are backfilled estimates.)
This line has no spike. It adds between +12 and +20 per day, very evenly, with no change around the rename. It never made GitHub Trending and never made the Trendshift top 25. I found no posts about it on X, Hacker News, or Reddit, and awesome-REA list 11-05 hasn't added it in updates. On PyPI, crab-cleanroom got 133 downloads last week, and the old name teardown-cleanroom got 25, though that includes mirrors and bots (pypistats, 11-06). It has 0 issues and 0 PRs, and I see no sign anyone has actually used it to clone anything.
Where did the stars come from? The exposure I could trace is all in automatically aggregated places: MCP directories like Glama, mcpjunction, M8ven, and MCP Radar; star-tracking sites like trendshift and upstar; PyPI; plus the 10 GitHub topics it filled in (clean-room-reimplementation, model-context-protocol, and so on).
The author matters more. From Oct 21 to 24, over four days, he shipped 4 agent-adjacent tools in a row:
| Repo | What it does | Stars (11-06) |
|---|---|---|
| veil | a privacy proxy for agents | About 120 |
| crab | a behavioral clean-room cloner | About 200 |
| sill | a native Windows bash | About 170 |
| airlock | a local, private coding agent | About 100 |
The four repos have nothing to do with each other, yet each gained 10–15 stars a day, with nearly parallel curves and fork ratios of 15–20%. Each was set up with the full set of distribution channels: PyPI, Homebrew, Scoop/winget, and then MCP directories. He has 146 followers, and his profile links the repos to each other (inferred). So CRAB's stars are half from the "compliant cloning" pitch and half from the author's distribution playbook, and I can't tell which contributed more.
Star farming can't be confirmed or ruled out from outside: the stargazer list isn't accessible, the curve is suspiciously smooth, and four repos gaining in lockstep is worth being skeptical about.
Who it's competing with
| Name | Scale | How CRAB differs |
|---|---|---|
| morluto/rea | 436k stars | Uses a decompiler to dig into binaries; one agent does everything end to end, with no clean-room process |
| aicodeking/leanrea | 4,019 star | A lightweight REA alternative focused on speed and cost; ignores isolation and verification |
| whit3rabbit/clean-room-skill | 19 star | Uses multiple agents to write specs, but starts from source code |
| devtechedge/feature-heist | 0 star | The closest in approach, with differential testing, but its input is REA's decompiled evidence |
| climashscape/agentic-cleanroom | 2 star | A methodology with no enforced isolation environment |
| Claude Code / Codex + Playwright MCP | — | Can assemble observations ad hoc, but has no gate, isolation, or audit |
The whole "AI reverse-engineering and cloning" category is already crowded, but in the small cell of "behavior only, enforced isolation, auditable," CRAB ranks first at two hundred stars, with everything else in single digits. That means almost nobody occupies this cell, and it doesn't hold many stars.
What's worth learning from it
Find where a big wave gets criticized most, and do the opposite. REA was criticized for being "one agent doing everything, not a real clean room," so CRAB built two agents, container isolation, and signed audits. This position doesn't go head-to-head with the wave, and it still rides the wave's traffic.
Publish results on a hidden test set in the README, including failures. In a category full of boasting, honesty is the differentiator.
One person ships several small tools in a week. Each goes to PyPI, Homebrew, and Scoop, gets listed in MCP directories, has its topics filled in, and links to the others from the profile. No single repo breaks out, but together they can steadily pull in a dozen or so stars a day.
Change the name early. "Teardown" is a generic word and collides with a game; CRAB is a short, memorable acronym that comes with a crab mascot.
If it were you
The verdict: C. Copying CRAB itself isn't worth it. Its cell tops out at a few hundred stars. REA's wave is still growing, but many projects are already riding it, and CRAB itself only got two hundred stars, half of which should be credited to the author's distribution push. What's worth taking is the playbook: take the opposite side of a controversy within a hot wave, then spread it out over a week of releases.
Who it fits: developers already building tools in an agent ecosystem who want stars and recognition, especially those who can ship several small libraries in a row and are willing to package each one for PyPI, Homebrew, and MCP directories.
Who it doesn't fit: anyone hoping a single repo will take off. CRAB's curve was built from GitHub's internal long-tail traffic, with no single external push. If you ship just one, it will likely stall at a few dozen stars. It's also a poor fit for anyone wanting to build a head-on REA competitor, since that space already has the leader at 430k stars and a lightweight version with 4k stars.
How to do it:
- Use the same move on the next wave: watch GitHub for agent tools gaining tens of thousands of stars a week, look through their issues, awesome lists, and critical blog posts for the most common complaint (unsafe, non-compliant, too expensive, not auditable), and build a small library that solves exactly that.
- If you insist on staying in the cloning space: fix what CRAB got wrong. It's too heavy, requiring Docker, Playwright, and terminal sign-offs, and 0 issues suggest almost nobody has actually run it. A 8KB "behavior-only" skill (like leanrea was for REA), or filling in the macOS, Windows, and iOS targets it lists as not yet started, has a better chance than building another all-in-one suite.
- Copy its README structure: open with a statement of position, follow with a flow diagram, then a single command that runs the demo, then hidden-test-set results with failure cases, and finally a clear list of what it will never do.
How to get users: within a week, release three or four small libraries on the same theme, each with fully filled-in topics, published to PyPI, Homebrew, and Scoop, submitted to MCP directories like Glama and mcpjunction, and cross-linked from the top of each profile. One step CRAB skipped that you can add: submit the repo to the awesome list for the wave it belongs to (CRAB isn't in awesome-REA), and leave links in the comments of the blog posts and threads that critiqued the controversy.
When to move: if riding REA's wave, the sooner the better, no later than the end of 2026. Once the wave recedes, nobody searches for "compliant cloning" anymore. For the next wave, launch within the first two weeks of its peak.
What would show this analysis is wrong: ① after late November, CRAB keeps gaining more than 20 stars a day without the author shipping new repos, or people start filing issues and writing tutorials, which would mean this cell is bigger than it looks and worth going after head-on; ② someone ships a repo built on the "opposite of the wave's controversy" playbook, gets listed on an awesome list, and still hasn't passed a hundred stars after a month, which would mean the approach depends mostly on the author's existing audience and can't be copied; ③ a model provider such as Anthropic explicitly restricts reverse-engineering-style cloning use, and the whole line shuts down.
Data and sources
GitHub repo page and daily star/fork counts (10-27 to 11-06), the author's profile and his 4 repos from the same period; the CRAB README (11-06); Trendshift daily counts; pypistats (crab-cleanroom, teardown-cleanroom, 11-06); the awesome-reverse-engineer-anything list (11-05); the redreamality.com blog on AI reverse-engineering and clean rooms; X, Hacker News, and Reddit searches found no relevant posts; listing pages on Glama, mcpjunction, M8ven, and upstar. The source of the initial ~50–60 stars, and whether stars were purchased, are unverified.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one morluto/rea Open source 2026-04-14 436k stars / 108k forks Uses a decompiler to pull apart binaries; one agent goes from analysis to reconstruction with no clean-room isolation. CRAB looks only at behavior, uses two separate agents, and enforces isolation. aicodeking/leanrea Open source 2026-10-17 4019 star A lightweight skill that replaces REA, pitched on speed and low cost; does not handle isolation or verification. devtechedge/feature-heist Open source 2026-10-21 0 star Turns REA's decompiled evidence into a clean-room TS rewrite of a single feature, with differential testing; its input is not pure black-box. whit3rabbit/clean-room-skill Open source 2026-05-20 19 star Starts from source code and uses multiple agents to write specs; CRAB never looks at source. climashscape/agentic-cleanroom Open source 2026-10-10 2 star A methodology plus a toolset, with no enforced isolation environment. websitebutlers/cleanroom-mcp Open source 2026-04-03 6 star A clean-room MCP with the same name and very small scale; features unverified. drevendev/HungryCrab Open source 2026-09-05 5 star / 14 fork Extracts code from open-source repos in a compliant way; the target is open-source code, not a black-box application. GhidraMCP / ida-pro-mcp Open source 2025-03 12k / 13k stars Pure decompilation-based reverse engineering; direction opposite to CRAB's. Assembled ad hoc with Claude Code / Codex + Playwright MCP. Big-company features — Very large Can make observations, but has no spec gate, isolation, audit, or cross-check verification. - Where users come from
- Verified Riding the REA wave: REA went from 37k to 430k stars in October, which spawned a crop of clean-room cloning projects and a legal controversy; CRAB launched 10-24, and its README directly responds to that controversy.
- Verified Automatically indexed by MCP directories and star trackers: Glama, mcpjunction, M8ven, MCP Radar, latgenhub, trendshift, upstar, and the repositorystats MCP topic trend page all list it.
- Inferred Cross-promotion among the author's concurrent repos / follower activity: veil, crab, sill, and airlock were released in succession from 10-21 to 10-24, each gaining 10–15 stars a day on nearly parallel curves; the author has 146 followers.
- Inferred GitHub topics and search: 10 topics were added, including clean-room-reimplementation and model-context-protocol; the curve is a steady long tail.
- Verified PyPI distribution: crab-cleanroom had 133 downloads in a week, teardown-cleanroom had 25.
- Inferred X / HN / Reddit posts: searches on x-radar, HN Algolia, Jina, and Brave turned up nothing.
Why lemonadesaltbagel/crab (CRAB, formerly Teardown) got traction
Channel: GitHub. Research date 2026-11-06. Each figure notes its source; anything I couldn't confirm is marked "unverified."
One line
CRAB is a Python MCP toolkit for clean-room cloning that works on behavior alone. An observer agent treats the target software (CLI, web, Android) as a black box and writes its behavior up as falsifiable specs. Those specs pass through a gate that strips copied text and real names, then go to a separate clean agent that can only see the spec and rewrites the software. The rewrite is tested against the original, and only abstract pass/fail results come back. Both sides can be locked in containers for enforced isolation, and every crossing is logged to a signed, hash-chained audit log.
Bottom line: it's a companion tool to the 2026 October wave of "reverse-engineer and clone closed-source software with AI agents" (the REA wave), and it's riding that wave's traffic. Its angle is that it directly answers the wave's biggest controversy: REA-style tools rely on decompilation, and the same agent does both the analysis and the rewrite, so it isn't a real clean room. CRAB never touches code. Isolation is enforced by containers, plus an audit log and scores on a held-out test set. The absolute numbers are still small (about 200 stars, roughly 130 PyPI downloads a week), and the growth curve is a flat line with no sign of any single external spike. The same author shipped 4 agent tool libraries in the same week, and their curves look nearly identical. That suggests the traffic is coming from GitHub discovery, MCP directory sites, and cross-links from the author's profile, not from the product catching on with people on its own.
1. Data
- Repo created 2026-10-22, first commit on 10-24 23:49 (GitHub API). As of 11-06 there are only 7 commits in total, each one a large batch. The author is the only contributor, with 0 issues and 0 PRs.
- Stars: 10-27 51 → 11-06 196 (our github repo_day data); 11-06 live 210, forks 35, watchers 26 (GitHub API). Last 7 days +110, single-day peak +20 (11-05).
- Trendshift daily numbers (trendshift.io/repositories/322191): 10-25 and 10-26 were each +0,10-27 and +3, then +12 to +20,13 a day for 148 days in total. It never made Trendshift's top 25 and never reached GitHub trending (a repo under 60 days old can't get there anyway, see repo notes).
- Converted, that's about 50–60 stars on launch day and the next day, then a very even straight line.
- PyPI (pypistats): crab-cleanroom had 133 downloads last week, teardown-cleanroom 25, including mirrors and bots. Real usage is very small.
- Rename: 10-28 was published to PyPI as teardown-cleanroom 0.3.0; 11-03 renamed to CRAB 0.4.0. The growth curve didn't change before or after the rename.
2. Background: the REA wave (why the timing matters)
- morluto/rea (Reverse Engineer Anything): an MCP server plus CLI that lets coding agents call Ghidra / IDA / JADX to inspect binaries, Electrons, APKs, and websites, then rebuild their functionality. Created 2026-04-14,10-09, it hit 37k stars; 10-13 it grew to 180k, and 11-06 it sits at 436k stars and 108k forks (our GitHub database + GitHub API). This is the biggest wave on GitHub in October.
- A flood of "clean-room clone" projects followed: storytold's WordCraft (17k stars, created 10-07), CADCraft, GridCraft, FigCraft and a dozen other Craft apps, occupysaas/zeroauthy (cloning Auth0), salescompel (cloning Salesforce), plus BIMCraft, ProWave and more (our github search for clean-room; most of the new October repos are these).
- An entire ecosystem has formed around REA: aicodeking/leanrea (4019 stars, created 10-17, a 8KB skill that replaces REA), AIshihzong/awesome-reverse-engineer-anything (489 stars), REA-Prompts, rea-lab's comparison experiments, and more.
- The controversy is CRAB's entry point. The awesome-REA list's "Legal and controversy" section says: "critics note that one agent doing both analysis and rebuilding does not meet the personnel separation of traditional clean-room development," and notes that REA's own SECURITY.md says REA is not a sandbox. Redreamality's blog argues "if you want clean-room discipline with agents, enforce isolation at the filesystem and network level, not in the prompt" (redreamality.com/blog/ai-reverse-engineering-clean-room-rewrite-licensing). Lexology has a piece, "The Clean Room Problem: AI Reverse Engineering and the Limits of Safe…" (lexology.com, date unverified). movez's Substack tutorial "How to Reverse Engineer Anything with Opus 5.5" (10-12) includes a step specifically called "Separate the two jobs."
- CRAB's README turns these objections into product features, point by point: the observer and implementer are two agents; isolation is "enforced, not just instructed"; it doesn't fetch code or decompile, and stack traces are stripped at collection time; the gate runs verbatim and near-verbatim similarity checks; the audit log is signed and hash-chained; and in strict mode a human has to sign off in the terminal on every release.
3. Competitors
| Name | Where | Launched | Scale | How CRAB differs |
|---|---|---|---|---|
| morluto/rea | Open source (npm rea-agents) | 2026-04-14 | 436k stars | Calls decompilers down to the binary; one agent goes from analysis to rebuild; no built-in clean-room process. CRAB works from behavior only, keeps the two agents separate, and enforces isolation. |
| aicodeking/leanrea | Open-source skills | 2026-10-17 | 4019 star | A lightweight skill that replaces REA, focused on being fast and cheap; no clean-room isolation and no verification. |
| devtechedge/feature-heist | Open source | 2026-10-21 | 0 star | Closest in idea: turns REA's evidence into per-feature clean-room rewrites in TypeScript, with differential testing. But its input is REA's decompilation evidence, not pure black box. |
| whit3rabbit/clean-room-skill | Open-source skills | 2026-05-20 | 19 star | Starts from source code and uses multiple agents to write specs. CRAB starts from behavior and never looks at source. |
| climashscape/agentic-cleanroom | Open source | 2026-10-10 | 2 star | A methodology plus tooling (specs → oracles → ratchets), with no enforced-isolation runtime. |
| websitebutlers/cleanroom-mcp | Open source MCP servers | 2026-04-03 | 6 star | Called "clean-room MCP." Very small in scale; features unverified. |
| drevendev/HungryCrab | Open-source Claude plugin | 2026-09-05 | 5 star / 14 fork | Mines other people's open-source repos for useful parts while staying within license terms. The target is open-source code, not a black-box program. |
| LaurieWired/GhidraMCP, mrexodia/ida-pro-mcp | Open source MCP servers | 2025-03 | 12k / 13k stars | Pure reverse-engineering tools, the opposite direction from CRAB. |
| Claude Code / Codex stitched together with Playwright MCP | Big-company features | — | Very large | Can do similar observation, but has no spec gate, isolation, audit, or held-out verification. |
| Thoughtworks / Martin Fowler, "From Black Box to Blueprint" | Methodology article | Not verified | — | Covers a black-box-to-blueprint methodology, not a tool. |
| arXiv, "Rebuild Dossier: Mechanically-Enforced Specs for Agentic App Rebuilds" | Paper | 2026-10-29 | — | Academic work on the same idea (mechanically enforced specs), showing the direction is being researched. |
Bottom line: "AI reverse-engineering and cloning" is already a red ocean, with REA alone at 40+ million stars. But the niche of "behavior-only, enforced-isolation, auditable clean-room cloning tools" is held only by a handful of small projects with single- to double-digit stars, and CRAB ranks first at 200 stars.
4. Where users came from (growth tactics)
- No external spike found:
- X: x-radar local corpus over the last ~30 days searched lemonadesaltbagel, crab-cleanroom, teardown-cleanroom, and Clean Rebuild of Anything; each returned 0 results.
- HN: Algolia returns nothing.
- Reddit: Jina and Brave both find no dedicated threads.
- The awesome-reverse-engineer-anything list (updated 11-05) doesn't include CRAB or Teardown.
- Whatever exposure exists comes from automated aggregator sites:
- MCP directories: Glama (teardown-observe), mcpjunction.ai, M8ven (rated C 74/100), MCP Radar (mcp.liqiwa.com), latgenhub.
- Star-tracking sites: trendshift, upstar.dev ("+97 stars this week"), and the repositorystats trend page for the model-context-protocol topic.
- Package indexes: PyPI, piwheels.
- Plus GitHub topic pages: clean-room-reimplementation, specification-generation, model-context-protocol and 10 other topics are all tagged thoroughly.
- The same author's "matrix": author Xintian (Austin) Huang, Sydney, bio reads "Exploring agentic post-training," with 146 followers. Between 10-21 and 10-24 (4 days) he released 4 agent-tool libraries:
- veil: privacy proxy for agents, 120 stars
- crab:196 star
- sill: native Windows bash, 168 stars
- airlock: local privacy coding agent, 102 stars
This 4 repos gain 10–15 stars a day, and their curves run almost parallel (our github repo_day). They also all sit around 15–20% forks. In early October he also shipped a string of Ricoh GR camera firmware patches (GR4-LUTSmith 39 stars, etc.). That suggests the traffic mostly comes from his own GitHub presence: follower activity, his profile, and the repos funneling people to each other. All of them also have distribution set up: PyPI, Homebrew (homebrew-veil), Scoop / winget (sill), plus listings in MCP directories.
- Makeup of the forks: I looked at 10 fork accounts. They're all "collector" accounts, several years old, with about 30 repos and single-digit followers, habitually forking popular AI repos (GitHub API). They aren't doing real derivative work, and they aren't plagiarizing.
- Star fraud can't be ruled out or confirmed: the stargazer list isn't accessible in this environment (the API returns 404, GraphQL returns 0, and other repos are the same), so the quality of the starring accounts can't be checked. The curve is an even straight line with no peaks, 4 repos grew in sync, and the followers include suspicious usernames like gitstargazers and MacKeepUS. These are signals to watch, but they aren't evidence.
5. Why it took off
- Timing: REA turned "agents reverse-engineering and cloning closed-source software" into GitHub's biggest topic in October (430k stars), and the legal and clean-room debate came with it. CRAB launched in 10-24, right as people were arguing over what counts as a compliant clone.
- Its positioning is the flip side of that controversy: the README's first line is "Rebuild it from how it behaves, never from its code", and it packages "two separate agents + container-enforced isolation + a gate that checks for duplicates + a signed audit log" as the product. For developers and companies worried about legal risk, who want to clone a competitor's feature, that is a story REA can't tell.
- The README has hard numbers: on apps the model never saw, scored on a held-out test set, the CLI 397/397, three web apps 129/130, 126/130, 125/127, Android 113/118; Flipt only got 19/55, and that weakness is stated openly; it also admits that tests on famous targets like jq really measure the model's memory. In a wave full of "I cloned Photoshop" hype, this kind of honest evaluation makes engineers more likely to star it.
- Distribution: I found no marketing push. Growth comes from MCP directories, GitHub topics, PyPI, and the author's other libraries from the same period cross-linking each other. A steady line of 10–20 stars a day fits this long-tail, organic-discovery pattern.
6. Differentiation, and whether that's why it broke out
The differentiation has three parts: it only collects behavior-layer data (accessibility trees, API shapes, exit codes) and never touches code; isolation is enforced by containers, not written into the prompt; and it has a gate that can be audited and verified against the original. The REA side has none of these, and other clean-room projects either start from source code or are just methodology.
This is why it can get stars, but it is only a small part of the reason: in the REA ecosystem, "compliant cloning" was the one narrative slot not yet taken, and CRAB took it. Still, its scale (200 stars) is orders of magnitude below the wave itself (REA 430k, leanrea 4 thousand). The same author's other 3 unrelated libraries follow the same curve, so the author's GitHub distribution matters at least as much as the product differentiation.
7. Visibility and copycats
- Visibility: few people know about it yet. I found no discussion on X, HN, or Reddit, awesome-REA lists haven't picked it up, and only automated aggregator sites have pages for it. Revenue: it's an open-source tool with no revenue and no paid version.
- Copycats: I haven't seen anyone copying it. HungryCrab has a sub-agent called crab-cleanroom-impl, but HungryCrab was created in September and crab is its own mascot, so that's a name collision, not a copy.
8. What to borrow
- Find the biggest controversy in a large wave, then build specifically "the opposite of that controversy". For example, REA was criticized for having the same agent both analyze and rewrite, so it doesn't count as clean-room; CRAB builds enforced isolation and auditing.
- Put held-out test-set results in the README, and write up the failed cases too (Flipt 19/55, and jq measures the model's memory). In a field where boasting is the norm, honesty is itself a differentiator.
- One person ships several small agent-side tools a week: each comes with PyPI / Homebrew / Scoop / MCP directory listings, complete topics, and cross-linking through the author's profile, and each library reliably gets a dozen-plus stars a day.
- Rename early: Teardown is a generic word and shares its name with a game; CRAB is an abbreviation people can remember.
9. Still unclear
- I can't get the stargazer breakdown, so I can't rule out star farming, nor tell where the first 50–60 stars came from (possibly the author's own circle, or follower activity).
- Whether any real users have cloned a product with it: 0 issues, and roughly 130 PyPI downloads a week; there's no sign anyone is actually using it.
- I couldn't find the author's X account (the GitHub profile doesn't list twitter); whether anyone has promoted it through Chinese channels (MCP Radar is a Chinese site) is unverified.
- Whether legal or platform risk could emerge (for example, a claim that Anthropic is restricting REA-style usage, sourced from a Facebook group post, unverified).
Sources
- GitHub API: repos/lemonadesaltbagel/crab, its commits / forks, users/lemonadesaltbagel and its repos and followers
- Our github database: repo_profile, repo_day (crab, sill, veil, airlock, morluto/rea, aicodeking/leanrea, and the clean-room search)
- trendshift.io/repositories/322191
- pypistats.org/api/packages/crab-cleanroom/recent
- github.com/AIshihzong/awesome-reverse-engineer-anything
- movez.substack.com/p/how-to-reverse-engineer-anything
- redreamality.com/blog/ai-reverse-engineering-clean-room-rewr…
- www.lexology.com/library/detail.aspx?g=3ef72a54-2b2f-4bd0-90…
- arxiv.org/html/2608.23616
- martinfowler.com/articles/black-box-to-blueprint.html
- Listing pages: glama.ai/mcp/servers/lemonadesaltbagel/teardown, mcpjunction.ai/servers/lemonadesaltbagel--crab, m8ven.ai/mcp/lemonadesaltbagel/teardown, upstar.dev/r/lemonadesaltbagel/crab
- x-radar search (last 30 days, 0 posts), HN Algolia search (0 results)