reverse-engineer server
The read-only MCP server and CLI behind the reverse-engineer skill. The skill is the agent-facing entry point; this server replaces the commands that are missing or fragile on a given platform (otool, strings, jq, go version -m, asar extraction) with TypeScript parsers, and answers in at most 8 KB with citable locations. Installing the skill does not install the server.
| Tool | What it returns |
|---|
triage | Format by magic bytes, architecture, size, SHA-256, markers (Go, Rust, Electron fuses, N-API, packers) and the next call |
electron_surface | Cached asar extraction, main, preloads and exposeInMainWorld keys, .node addons and the addon names called from JS, IPC channels grouped by call site (app wrappers included), fuses, asar integrity |
search | Bounded search in a tree, a .asar or one file (hidden folders included), with file:line:col or file@offset and a short excerpt |
read | Line ranges, a column window for minified code (instead of reformatting it), hexdumps |
binary_inventory | PE, ELF and Mach-O sections, libraries and imports, exports, Node addon names, ASCII and UTF-16 strings, PE resources |
buildinfo | Go build info (toolchain, main package, module, settings, dependencies) and Rust cargo-auditable packages |
decompile | Ghidra headless, cached per file: list functions, map a string to the functions that use it, return one function's C |
managed | .NET (via ilspycmd) and APK/DEX/JAR (via jadx): decompile once, summarize, return one type |
The catalog is about 6 KB for the 8 tools; the skill is about 5.5 KB.
Design rules:
- Compact by default. Each answer stays under
REVERSE_ENGINEER_MAX_BYTES (8192). Longer output is written to the cache, and the answer ends with its path, so read and search can page through it.
- Cited facts. Every location is
file:line:col, file@0xoffset, a symbol or a string.
- No blocking on common cases. A
.node file signed after packaging shows up as an unverified hash, not as an error. Malformed or hostile binaries degrade to partial results instead of hanging.
- Read-only. The tools never run the target. They write only to their cache:
%LOCALAPPDATA%\reverse-engineer\cache, ~/Library/Caches/reverse-engineer, or $XDG_CACHE_HOME/reverse-engineer (override with REVERSE_ENGINEER_CACHE). Exceptions: Ghidra refuses project paths with a dot folder, so with such a cache path (the Linux default ~/.cache) its projects go to a private per-user folder in the temp directory; when the cache path contains a space, Ghidra keeps its settings in the home folder; jadx and the .NET SDK may write their own settings there on first use.
- Cache by SHA-256. An extraction or an analysis is reused.
Use
Requirements: Node.js 24.11 or later. Optional: Ghidra 11+ with JDK 21 for decompile, ilspycmd (.NET 10 SDK) and jadx (JDK 21) for managed. The agent can install them with the skill's install guide: official releases go into a per-user tools root (%LOCALAPPDATA%\reverse-engineer\tools, ~/Library/Application Support/reverse-engineer/tools, or $XDG_DATA_HOME/reverse-engineer/tools; override with REVERSE_ENGINEER_TOOLS), without admin rights or a PATH change. The tools look there first, then on PATH, GHIDRA_INSTALL_DIR and the usual install folders.
The package is published to npm as an alpha prerelease, reverse-engineer, and listed in the MCP Registry as io.github.Firzus/reverse-engineer. Pin the exact version:
npx -y reverse-engineer@0.1.0-alpha.2 triage ./some.exe
npx -y reverse-engineer@0.1.0-alpha.2 electron_surface "C:\Program Files\App" --focus audio,microphone
Claude Code MCP configuration (project .mcp.json):
{
"mcpServers": {
"reverse-engineer": {
"command": "npx",
"args": ["-y", "reverse-engineer@0.1.0-alpha.2", "mcp"]
}
}
}
On Windows, use "command": "cmd" with "args": ["/c", "npx", "-y", "reverse-engineer@0.1.0-alpha.2", "mcp"]. To run an unreleased build from a clone, run the Develop commands, then point the configuration at node <checkout>/mcp/reverse-engineer/dist/cli.mjs mcp.
Install the skill with npx skills add Firzus/agent-skills --skill reverse-engineer.
Develop
The toolchain is Vite+ (vite-plus, local dev dependency). .github/workflows/reverse-engineer-check.yml runs the same checks on Windows, macOS and Linux:
npm install
npx vp check
npx vp test
npx vp pack
Tests against real targets are opt-in: set RE_TARGETS (a folder with app.asar, app.asar.unpacked/ and gh.exe), RE_NOTION_DIR (an installed Notion app), RE_GHIDRA=1, RE_DOTNET_DLL, or RE_APK.
Release
.github/workflows/reverse-engineer-publish.yml publishes to npm and the MCP Registry. Merging a version change to main publishes it, so treat the merge as the release decision. In the release pull request:
- Set the same version in
package.json, package-lock.json, server.json, the pinned commands above and in the skill's SKILL.md; a unit test checks the first three. When vp pack reports a new bundled dependency, add its license to THIRD_PARTY_NOTICES.md.
- Run the Develop checks, then
npm pack (it rebuilds dist/) and inspect the tarball: LICENSE, README.md, THIRD_PARTY_NOTICES.md, dist/cli.mjs, ghidra/ExportDecompiled.java, package.json.
The workflow stops when the version is already on npm and the MCP Registry. Otherwise it runs the check workflow, builds the tarball in a job without publishing rights, then publishes in the release environment, from a job that runs no project or dependency code. npm uses trusted publishing, with a prerelease published under its own dist-tag (alpha), never latest; the MCP Registry uses its GitHub OIDC login. A last job runs the published --version. When a run fails, fix the cause and use Re-run all jobs on that run: each publication step rechecks its registry and skips what already exists. Never overwrite or unpublish a version to hide a failed release; publish a new one.
npm must trust the workflow before the first automated release: in the reverse-engineer package settings, add a GitHub Actions trusted publisher with user Firzus, repository agent-skills, workflow reverse-engineer-publish.yml and environment release.
Benchmark
bench/ adapts the harness from leanrea to Windows and compares four arms: no skill, leanrea, REA (MCP and skill), and this tool. See bench/BENCHMARK.md for the method, the results and their limits.
License
MIT. The benchmark tasks and harness are adapted from leanrea, MIT-licensed; its notice is in bench/LICENSE-leanrea. Fuse names and the Go build-info layout follow the public formats documented by Electron and the Go toolchain; no code is copied from REA.