REA started as a tool for reading closed software. It is now a whole genre of projects. Since the repository passed hundreds of thousands of stars, people have pointed it at games, screen recorders, design apps and desktop clients, then published what they rebuilt. We checked GitHub directly and picked ten public projects that credit REA as part of how they were made.
If you have not met the tool yet, start with our REA setup and security explainer and the case-study breakdown. This post is different: it asks what do people actually build with the REA skill, and which results hold up?
TL;DR: the ten projects at a glance
The list is ordered by GitHub stars on October 19, 2026. That is a measure of attention, not quality, so each entry also says what the evidence is.
| # | Project | What it is | Stars | License | Evidence level |
|---|---|---|---|---|---|
| 1 | Open Claude UI | React UI kit rebuilt from a desktop AI client | 1,217 | MIT | Author claims; no independent check |
| 2 | Co-op Anything | Agent playbook that adds co-op to single-player games | 437 | MIT | Playbook recommends REA; games ship separately |
| 3 | rec-rs | Rust macOS screen recorder, OBS alternative | 83 | Apache-2.0 | Architecture study |
| 4 | REA Reconstruction Companion | Agent skill that plans and validates rebuilds | 79 | MIT | Tool, not a rebuild |
| 5 | DX-Ball reconstruction | C/C++ rebuild of a 1996 game | 66 | MIT | Strong: differential tests |
| 6 | paper-rs | Rust infinite-canvas design app | 54 | Apache-2.0 | Behavior study |
| 7 | TH04 | Source reconstruction of a PC-98 game | 34 | MIT | Strong: replay comparisons |
| 8 | Open Scrantic | Rust engine for Johnny Castaway | 17 | MIT | Public CI, in progress |
| 9 | Snowball War | Rust/SDL2 rebuild of a Ski or Die minigame | 2 | Unspecified | Evidence-driven spec |
| 10 | hexgrab | .NET color picker from PowerToys study | 1 | MIT | Best-documented accuracy check |
What is the REA skill, in one paragraph?
REA is an MCP server and CLI that connects your agent to Ghidra, Hopper, IDA and static analyzers for JavaScript and .NET. The reverse-engineer-anything skill is the instruction file that teaches the agent the workflow: inspect, cite evidence, list unknowns, then rebuild. If you want the background on how skills work, read our agent skills guide and the MCP explainer.
Setup is npx rea-agents setup. Community docs also cite a skill-only route, npx skills add morluto/rea --skill reverse-engineer-anything. Node 22.19 or newer is required, and native work needs Hopper, Ghidra or IDA. Our Claude Code commands reference covers the agent side.
1. Open Claude UI: a Claude-style component kit
Open Claude UI is the most-starred project here at about 1,217 stars. Its README says the author used REA's static AST parsing, deobfuscation and ASAR unpacking on a modern desktop AI client to extract component specs and animation constants. The description cites 30 independent React components, 8 interaction-physics algorithms and a manual of more than 3,000 lines. It is MIT-licensed TypeScript with a badge reading "Reverse Engineered with REA."
What to watch. The numbers come from the author, not from a test suite comparing against the original. The repository also carries an asset-compliance note about clean-room assets, which is the right instinct: a look-alike of a commercial product's design raises trademark and licensing questions that code analysis does not answer.
2. Co-op Anything: turn single-player games into co-op
Co-op Anything (about 437 stars, MIT, created October 16) is a playbook rather than a game. You tell your agent to read the repo and follow its AGENTS.md, and the agent researches the game, reverse engineers it if needed, builds the co-op, tests it with bots and packages it. The README lists games in progress including Deus Ex (released), Fallout: New Vegas and Zelda: Ocarina of Time (fully synced).
What to watch. REA is a recommended companion here, not a dependency. The README says to pair the playbook with a reverse-engineering MCP such as REA: "Co-op Anything knows what to do, REA looks inside the game." The individual games live in separate repositories, so judge each one on its own.
A game controller with a green spark, representing a single-player game rebuilt for co-op with REA
3. rec-rs: a Rust screen recorder studied from OBS
rec-rs is a fast macOS screen and camera recorder in Rust, described as a clean-room alternative reverse engineered from OBS Studio's architecture. About 83 stars, Apache-2.0.
What to watch. OBS Studio is GPL-2.0. Studying an open-source program's architecture is allowed, but a permissively licensed rewrite should be careful not to copy code. The author's own note describes it as a clean-room effort, so read the repo's provenance notes before reusing anything.
4. REA Reconstruction Companion: a skill for the rebuild half
REA-Reconstruction-Companion is itself an agent skill. REA gathers evidence; this companion turns it into requirements, architecture, a staged roadmap and validation steps. It is model-independent, MIT-licensed, and has about 79 stars. Its table maps goals to outputs, such as "Investigate a feature, function or format" to a focused contract plus acceptance cases.
Why it matters. It targets the weakest step in most REA projects: going from findings to a plan you can test. Its integration with REA is optional and explained before install. Pair it with the Replica skills guide if you want to compare approaches to cloning app behavior.
5. DX-Ball reconstruction: the best-tested game rebuild
N0zoM1z0/dx-ball rebuilds English DX-Ball v1.07 as readable C/C++. The README describes a loop: inspect a function through REA and Ghidra, recover typed source, then compile with the original Visual C++ 4.0 compiler under Wine and compare the machine code to the original. About 66 stars, MIT.
This is the project behind REA's showcase result: the sound-pan function matched 3,205 original x86 cases. It is also a first-party case, since the author is a REA contributor. The README is candid that exact reconstruction is still in progress and the game is not yet playable.
A path ending at a flag, a milestone for a game rebuilt with the REA reverse-engineering skill
6. paper-rs: a design canvas studied from behavior
paper-rs is a native Rust (egui and wgpu) infinite-canvas design app. Its README says it was built by studying the Paper desktop app's behavior with REA and Claude Code, and states that no original code or assets are included and that it is not affiliated with Paper. About 54 stars, Apache-2.0, with an MCP server so agents can read and design on the open canvas.
What to watch. "Behavior study" means the author observed what the app does. It is not a claim of binary-level equivalence.
7. TH04: the PC-98 reconstruction behind a showcase
N0zoM1z0/th04 reconstructs Touhou 4: Lotus Land Story for PC-98. Its README reports detailed replay comparisons, for example 13,378 pool and state snapshots and 22 full RGB samples matching a recorded stage replay, while listing differences it has not closed. About 34 stars, MIT.
What to watch. The README itself does not name REA; the connection comes from REA's official showcase page. Treat it as a related case, and note the project's own honest status: complete game and timing acceptance remain unfinished.
8. Open Scrantic: Johnny Castaway in Rust
Open Scrantic (about 17 stars, MIT) rebuilds the engine for the Windows 3.1 screensaver Johnny Castaway in Rust for Windows, macOS and Linux. Per a curated case log, the author said REA handled reverse engineering, Claude Code did implementation, and they reviewed the work. The architecture splits resource decoding, story planning, script scheduling, rendering and sound.
What to watch. You must supply your own original game files. The project documents deliberate deviations, such as full-frame redraw instead of dirty rectangles, so "same behavior" is not a claim of frame-exact identity.
9. Snowball War: one minigame, one spec
rea-ski-or-die-reconstruction rebuilds a single minigame from the 1990 DOS game Ski or Die in Rust and SDL2. Its README says REA analyzes the shipped binary, recovers gameplay rules with evidence, records them, and rebuilds from that spec. Only Snowball War is in scope. About 2 stars; no license identified on GitHub.
Why it is useful. It shows the "spec first" pattern at the smallest practical size. You download the original DOS release you own and build assets yourself.
10. hexgrab: the clearest accuracy check
hexgrab is a color picker under 750 lines of C#. The author used REA on the shipped binaries of Microsoft PowerToys' Color Picker, then checked each finding against the public source. Per the README, REA got 8 of 9 findings fully right, documented in docs/rea-findings.md. About 1 star, MIT.
Why it deserves a slot. It is the rare project that measures the tool. Because PowerToys is MIT-licensed, the author could verify every claim. The README also warns people not to point reverse-engineering tools at commercial software they have no right to study.
What do the best projects have in common?
Read the ten side by side and a pattern appears.
| Habit | Who does it | Why it matters |
|---|---|---|
| Narrow target | Snowball War, hexgrab, DX-Ball | A function or minigame can be checked; a whole app cannot |
| Tests against the original | DX-Ball, TH04 | Replay and differential tests catch confident mistakes |
| Verify against known source | hexgrab | Measures how often the agent is right |
| Name the limits | Open Scrantic, DX-Ball | Honest status builds trust |
| Respect licenses | hexgrab, paper-rs | Avoids copying code or assets |
The weakest results share a shape too: large claims, no comparison against the original, and a commercial product on the other end.
What people are asking
Is a high star count proof it works? No. Open Claude UI has far more stars than DX-Ball, yet DX-Ball publishes the stronger evidence. For more on separating claims from checks, see how AI agents decompile games, and the browser ports of decompiled games.
Will companies tolerate this? Mixed. Big studios are exploring it themselves, as in Xbox Labs' game decompiling and recompiling work, while others push back. Our PhotoCraft coverage shows the clean-room argument in action.
Can I copy these projects? Check each license. Several require original game files, and none of them grants rights to the studied software.
Try it yourself: a safe first prompt
Pick something you own or an MIT-licensed program, then ask a narrow question:
Using REA, study how this open-source color picker finds the pixel under the
cursor. Show the evidence for each claim, list what you could not confirm,
then write a minimal version and a test that compares it with the original
on at least 100 inputs. Do not copy any decompiled code.
This copies hexgrab's method: study, verify against known source, rebuild small. If you want guided practice with agent workflows, see our workshops.
Limits to keep in mind
- Studied behavior is not rights to copy. A faithful rebuild can still infringe if it reuses assets or trademarks.
- Star counts move fast. Figures here are from October 19, 2026 GitHub data.
- Some credit is secondhand. TH04's README does not mention REA, and Co-op Anything only recommends it.
- Early projects change. Many repositories were created in the past two weeks and will look different soon.
Related reading
- REA MCP server: what agent reverse engineering changes in your setup
- Reverse engineering is now a prompt: REA's DX-Ball, Notion and TH04 case studies
- How AI agents decompile games: workflow, limits, claimed vs verified
- AI-decompiled games in the browser
- What are agent skills? A complete guide
- Replica skill: clone any app
- PhotoCraft and the clean-room question
- How to install and use the REA skill with Claude Code
- Prove an AI reverse engineering result with tests
- Legal and clean-room checklist for AI reverse engineering
- Official: REA on GitHub · REA showcase
Star counts, licenses and README details were checked on GitHub on October 19, 2026. These repositories are new and change often; verify current status before relying on any of them.