— reading now
Flash

TopicTrending open-source AI on GitHub

Review 8.6·Apps & Agents·2026-10-13 22:35

REA Review: Letting Agents Reverse-Engineer Anything, 100K Stars in a Week

Open-source REA gained 100K stars in a week: an MCP server that lets coding agents like Claude Code, Codex and Cursor call decompilers such as Ghidra, IDA and Hopper directly — reading software without its source.

Decompiled code on screen with the REA headline
REA cover

Scorecard

8.6/ 10

HIMIAO Horizon score

Infrastructure-grade tooling: testing proves the static analysis works; deep native analysis depends on backend setup.

Pros

  • Broad MCP ecosystem support

  • Static JS analysis verified fast and solid by independent testing

  • MIT-licensed with fast iteration

Cons

  • Independent testing found misses

  • Native analysis needs paid or heavy backends

  • Legal boundaries are the user's own responsibility — analyze only software you are allowed to

Best for

Developers who need to understand closed-source apps, build compatible implementations or play CTFs

01What happened

REA (Reverse Engineer Anything, morluto/rea, MIT-licensed) began in April 2026 as betterBinaryMCP. Version 4.1.0 on October 6 added JADX Android APK analysis, Binwalk/Unblob firmware analysis and read-only IDA access, followed by rapid 5.0.0 (October 7) and 6.0.0 (October 8) releases. Stars climbed from under 10K on October 6 to over 38K by October 9 and over 100K by October 12 (one report cites 115K), topping GitHub's daily trending with 25K+ new stars in a single day. How it works: an agent calls REA over MCP, and REA routes the question to a local decompiler or disassembler (Ghidra, Hopper, IDA, JADX, firmware tools), returning pseudocode, assembly, strings and “evidence and limitations.” Installation: npx rea-agents setup.

02Testing and results

The basis is a published independent hands-on test (dev.to, October 12, 2026): a Debian 13 box with 8 vCPUs, no GPU and no cloud API keys; REA was driven through its own CLI and a hand-written MCP script (testing the tools, not an agent's judgment); targets were a 100MB Electron app (459 files) and a self-built 5-file toy app with known answers. Results: static JS analysis was fast and solid, but missed two preload files visible via grep, and the tooling places a heavy call burden on agents. Vendor case study: one DX-Ball function rebuilt in C passes 3,205 test cases against the original x86 code (vendor data). For context: SRE-Bench shows the best agent fully solving 31.5% of realistic test binaries.

03Key facts

  1. Works with: MCP agents including Claude Code, Codex, Cursor and Gemini CLI.
  2. Backends: Ghidra, JADX and firmware tools are free; Hopper is paid; IDA needs the Pro edition; Ghidra and IDA paths are read-only analysis.
  3. Project health: CI, a security policy, a docs site and a Discord community; the README ships in Chinese, Japanese, Korean and Arabic.
  4. Disputes: Activision reportedly pulled AI-decompiled game footage, and Bethesda issued a cease-and-desist against an AI-generated browser game.

Why it matters

100K stars in a week shows real developer demand for agents that can read closed-source software — REA turned decompilation into an MCP tool agents can call.
Useful Tap if this story helped you

From HIMIAO Horizon · himiao.ai

Comments

  1. Loading comments…
Ask the cat