Binary analysis · APK & IPA decompilation · .NET / Java · Electron & WASM · firmware · malware triage · API & protocol recovery · Ghidra / IDA / radare2 / jadx via MCP

Claude Reverse Engineering Skill

An Agent Skill that gives your AI agent a professional reverse-engineering workflow: identify what a file really is, pick tools that are actually installed, decompile and export code headlessly, map the network surface, and report findings that point to real addresses — while handling untrusted samples safely.

It covers native executables (ELF, PE, Mach-O), Android (APK, AAB, DEX), iOS and macOS apps, .NET and Java bytecode, JavaScript/Electron/WebAssembly, cross-platform frameworks (Flutter, React Native, Xamarin/MAUI, Unity), installers, firmware images, and network protocols.

Built for authorized work: security research, CTFs, malware analysis, interoperability, and debugging.

Contents

• Why this skill
• What it can analyze
• Works with
• Install
• Usage
• Bundled tools
• How it works
• Requirements
• What the skill runs
• FAQ
• Responsible use
• Contributing
• License

Why this skill

• One command to install, nothing to configure. A standard Agent Skill: install it as a Claude Code plugin or with npx skills add, and it works. No cloning, no index scripts, no per-client adapters, no extra runtimes — just Python 3.8+.
• Scripts that do the analysis. The bundled tools read the files themselves — triage, app/platform metadata, endpoint extraction, decompiled-code export — instead of only routing the agent to someone else's tool.
• Triage first, not guesswork. A zero-dependency triage script reads magic bytes and real headers (PE, ELF, Mach-O, CLR), fingerprints 25+ frameworks, packers, and installers, and tells the agent exactly which playbook to open next.
• Headless by design. Agents can't click through a GUI. The skill ships a Ghidra post-script that exports decompiled C, imports, and strings to a text file — tested against Ghidra 12.1.4, with a CI job that re-runs it on every push — plus non-interactive commands for rizin, jadx, Frida, gdb, and lldb.
• Drives real RE tools over MCP. When rea, GhidraMCP, IDA Pro MCP, radare2-mcp, or jadx MCP is connected, the agent decompiles, renames, and follows cross-references directly — the skill supplies the method, the server does the calls.
• Finds the network surface fast. Pulls Retrofit annotations, HTTP client calls, URLs, hosts, auth headers, and crypto usage out of decompiled code and minified JS, each with file:line.
• Safe with untrusted samples. Isolation rules for malware, python -I against import hijacking, and an explicit rule that text inside a binary is data — never instructions — so analyzed files can't prompt-inject the agent.
• Grounded output. Every finding must cite an address, offset, symbol, or request the tools actually showed, say how it's known (decompiler, assembly, runtime, or inferred), and name what wasn't verified.
• Tested on real tools, not just docs. 63 automated checks covering every script and the skill's structure, run in CI on Linux, macOS, and Windows, plus a CI job that runs the Ghidra export against a real Ghidra release on every push.
• Light on context. One short SKILL.md; the 14 playbooks load only when a task needs them, so the skill costs almost nothing until it's used.
• One permissive license. Everything is Apache-2.0 — safe to use at work, fork, and ship.

What it can analyze

Works with

Any agent that can read a SKILL.md and run shell commands can use the full workflow. Chat-only assistants without a shell can still follow the playbooks, but can't run the bundled scripts.

Install

Claude Code (recommended)

Listed in Anthropic's official plugin directory for Claude Code, Cowork and the Claude apps: search for Reverse Engineering there. Or install it as a plugin straight from this repository — updates arrive automatically. Inside Claude Code, run:

/plugin marketplace add EncidXVII/claude-reverse-engineering-skill
/plugin install reverse-engineering@reverse-engineering-skill

Or from your terminal:

claude plugin marketplace add EncidXVII/claude-reverse-engineering-skill
claude plugin install reverse-engineering@reverse-engineering-skill

Or copy the skill folder (personal = all projects, project = this repo only):

git clone https://github.com/EncidXVII/claude-reverse-engineering-skill
cp -r claude-reverse-engineering-skill/skills/reverse-engineering ~/.claude/skills/        # personal
cp -r claude-reverse-engineering-skill/skills/reverse-engineering .claude/skills/          # project

Claude.ai and Claude Desktop

Add Reverse Engineering from Anthropic's plugin directory. Or download the repository, zip the skills/reverse-engineering folder, and upload it under Settings → Capabilities → Skills.

Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, Windsurf, Kiro

Use the open-source skills installer (the skill is listed on skills.sh):

npx skills add EncidXVII/claude-reverse-engineering-skill --skill reverse-engineering
npx skills add EncidXVII/claude-reverse-engineering-skill --skill reverse-engineering -a codex -a cursor   # pick agents
npx skills add EncidXVII/claude-reverse-engineering-skill --skill reverse-engineering -g                   # user-wide

Copy skills/reverse-engineering into the agent's skills directory:

Usage

Just describe the task — the skill loads automatically. No special syntax needed.

I have a stripped ELF from a CTF (./crackme). What does the password check do?

Decompile base.apk and list every API endpoint it calls, with the auth headers.

What is this installer from our vendor actually going to drop on the machine?

Unpack this router firmware and find the web admin backend.

Here's a malware sample in my analysis VM — summarize its behavior and IOCs.

How does this Electron app talk to its native addon?

In Claude Code you can also invoke it directly with /reverse-engineering (or /reverse-engineering:reverse-engineering when installed as a plugin).

Bundled tools

All scripts are standard-library Python — nothing to pip install.

Real triage.py output for a Rust executable (trimmed):

type        : PE (Windows .exe/.dll/.sys)
format      : PE32+
arch        : x86-64
subsystem   : Windows console
dotnet      : False
entropy     : 6.428/8.0  (moderate — typical of native code with data)
SECTIONS
  .text      raw  9,209,856  entropy 6.31
  .rdata     raw  3,086,336  entropy 6.03
FINGERPRINTS
  * Rust binary — demangle names (rustfilt), expect heavy inlining; panic strings leak source paths
SUGGESTED NEXT STEPS
  - references/native-binaries.md — `rz-bin -I` / `rz-bin -i` (imports), then decompile
    headless with scripts/ghidra/DecompileExport.java (SKILL.md step 6)

Run them yourself:

python skills/reverse-engineering/scripts/check_tools.py
python -I skills/reverse-engineering/scripts/triage.py path/to/file-or-folder
python -I skills/reverse-engineering/scripts/extract_endpoints.py path/to/decompiled-output

How it works

1. Toolbox check — find out what's installed before planning.
2. Triage — identify the artifact from its bytes, not its extension.
3. Safe workspace — one sample per empty folder; nothing untrusted runs on the host.
4. MCP first — drive Ghidra/IDA/radare2/jadx directly when a server is connected.
5. Route — open the playbook for the target type.
6. Static first, dynamic to confirm — headless decompilation, then hooks or a debugger in isolation when values only exist at runtime.
7. Record — findings with addresses and offsets, using the report template.

Requirements

• Required: Python 3.8+ and an agent with shell access.
• Optional, detected at runtime: Ghidra 12+ with JDK 21, rizin or radare2, jadx, apktool, ILSpy (ilspycmd), Frida, mitmproxy, binwalk, Node.js. Install only what your target needs — check_tools.py prints the exact command for anything missing.

What the skill runs

Transparency for anyone reviewing it before install:

• No network access, no telemetry. None of the bundled scripts open a network connection or send data anywhere.
• Never executes the file being analyzed. triage.py, app_inspect.py and extract_endpoints.py only read bytes; the Ghidra scripts run inside your own Ghidra.
• check_tools.py looks up tool names on your PATH and in common install folders (or a --ghidra-dir you pass), checks whether a few Python libraries are importable, and runs java -version.
• Everything else is guidance. The playbooks tell the agent which tools to run (for example Ghidra, jadx, or Frida) when you have them installed; your agent's permission settings decide whether it asks before running them.

FAQ

Can Claude reverse engineer an APK?
Yes. The skill has Claude decompile the APK with jadx, detect frameworks like Flutter or React Native, and extract the app's API endpoints with extract_endpoints.py.

Does it work with Ghidra MCP or IDA Pro MCP?
Yes. If an RE MCP server is connected, the skill tells the agent to use it for decompiling, renaming, and cross-references. Without one, it uses the bundled headless Ghidra export.

Does it work without Claude?
Yes. It follows the open Agent Skills standard and installs into Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and others with npx skills add.

Can it decompile any program?
It can analyze nearly any format, but results depend on the installed tools and on the target's protections: strong encryption without the key, or heavy virtualization-based obfuscation, can't be undone by any tool.

Is it safe to point at malware?
The skill never executes samples itself and enforces isolation rules for any dynamic step — but run malware analysis inside a disposable VM with no credentials, as you would without an AI agent.

Is reverse engineering legal?
It's widely legal for security research, interoperability, and analyzing software you own or are authorized to test, but laws and license terms vary by country. See Responsible use.

Responsible use

This project is for analyzing software you own, are authorized to test, or are permitted to study: CTF challenges, lab malware samples, your own products, and scoped security engagements. The skill declines to help with piracy, license or DRM removal to avoid paying, cheats for online multiplayer games, and making malware evade detection. You are responsible for following applicable laws and license agreements.

Contributing

Issues and pull requests are welcome — new target types, better tool recommendations, and fixes as tools evolve.

• Keep SKILL.md short (routing and principles); put depth in references/.
• Bundled scripts stay standard-library only and cross-platform.
• Every command in a playbook must be non-interactive.
• Run the test suite before opening a PR:

python tests/test_scripts.py

.
├── .claude-plugin/          Claude Code plugin + marketplace manifests
├── .github/workflows/       CI: tests on Linux/macOS/Windows, frontmatter lint, real-Ghidra export
├── evals/                   trigger evals for tuning the skill description
├── skills/reverse-engineering/
│   ├── SKILL.md             entry point: workflow, routing, safety rules
│   ├── references/          14 playbooks, loaded on demand
│   ├── scripts/             triage, tool check, endpoint extraction, Ghidra export
│   └── assets/              findings report template
└── tests/                   dependency-free test suite

License

Apache 2.0

_{Independent project, not affiliated with or endorsed by Anthropic, OpenAI, Google, GitHub, or the NSA. Claude, Ghidra, IDA Pro, and other product names are trademarks of their respective owners.}