reverse-skill

Why it's ranked here

The value here is discipline, not code. Most agent security packs are prompt piles. This one puts its routing decisions in one checked configuration, backs them with a regression suite that runs on two operating systems in continuous integration, and writes a hard consent gate into its rules: reading the repository is not permission to run it, and naming a target is not permission to attack it. That gate is rare in this category, and it is the reason to prefer this over an unstructured prompt collection. What it does not give you is the scanners and decompilers themselves. Those stay external, and you install them.

What's good

Three things stand out. The consent and scope model is explicit and repeated across the rule files: written authorization and a network profile must be on record before any target action, and a force flag is documented as unable to bypass that gate. The routing core is one structured table with a regression benchmark and cross-platform continuous integration, so a rule change that breaks matching fails the build. The proxy-control bridge ships with token authentication on by default and cross-origin access narrowed to the local machine, returning a rejection when the token is missing. Documentation is thorough and bilingual.

Tradeoffs

The counts do not line up. The landing page claims one number of routing rules and benchmark cases, while the file listing and the test runner cite lower figures, so the headline numbers do not agree. Almost all the weight is guidance, not executable capability: the scanners, decompilers, and the paid disassembler are installed by you, and one is commercial with no automated path. The scripts and most tooling assume Windows or Kali first, with generic Linux and macOS marked only as generic support. Continuous integration covers two operating systems, not the full matrix the docs describe. A large security payload corpus may be quarantined by antivirus.

How to use it well

Reach for this if you run a coding assistant on authorized reverse-engineering, contest, or permitted penetration work and you want the model to follow one repeatable path instead of guessing at tools. It fits teams that already have the underlying scanners and decompilers installed and want consistent routing plus evidence logging on top. It does not replace those tools, teach the tradecraft, or help with unauthorized targets: the rules block that by design. Read the on-demand installers closely before approving them, since they fetch and run external software. Solo users on macOS get the least polished path.

Observed

License

MIT

Primary interfaces

Agent skill-router instruction files plus an MCP server bridging a web proxy tool

Distribution

Clone the Git repository; no published package-manager entry

Runtime prerequisites

Java (JDK), Node.js 22.12+, Python 3.x

Platform support

Windows and Kali as primary paths, Ubuntu/Debian and macOS as generic

Routing tests

In-repository routing regression suite, run in CI on Windows and Ubuntu

Consent model

Rules require recorded authorization and a network profile before any target action

Proxy bridge auth

Bearer token required by default; cross-origin access limited to the local host