Hand a coding agent an APK to analyze, a binary to reverse, or a pentest scope to work through, and it faces the same problem a junior analyst does: which tool actually fits this job — jadx or apktool for this APK, Frida or IDA for this binary, and what's the actual repeatable methodology rather than guessed commands. reverse-skill is an Agent Skill built specifically to answer that routing question for reverse engineering, penetration testing, and CTF work — and, more interestingly, it builds authorization scoping and an audit trail directly into the routing logic instead of treating them as separate concerns.
What It's Actually For
The project is explicit about its scope in its own description: reverse engineering, authorized penetration testing, and security research — not a general hacking toolkit, and it says so upfront. It targets the specific pain points that show up once an AI coding agent is expected to work these problems: agents don't inherently know when to reach for jadx versus apktool versus Frida versus BurpSuite, an APK task and a PCAP task need entirely different playbooks, tooling and MCP servers end up scattered across machines, and without a system for it, the same mistakes get repeated because nothing captures what was learned last time.
The routing pipeline is described plainly in the project's own docs: a task comes in, gets checked against global rules, goes through a scope-and-authorization gate before any target is touched, gets routed to the right scenario skill and toolchain, and results in a timeline plus an Evidence → Finding → Path chain that becomes the actual report.
The Scope Gate Is the Part Worth Understanding
This is the detail that separates a routing pack built by people who've actually done authorized security work from a collection of tool wrappers: nothing acts against a target until a case has been initialized with authorization and a network profile confirmed. That's not a disclaimer bolted onto a README — it's a structural gate in the routing pipeline itself, described as "no target ACT until ready." Paired with a read-only case-review system that checks scope, timeline, work items, findings, and paths — with optional SHA-256 integrity checks on the evidence — the design mirrors how a professional pentesting engagement is actually run: authorization confirmed first, every action traceable back to evidence, and a reviewable chain from finding to report rather than a loose collection of terminal output.
What's Actually Covered
The routing matrix spans a genuinely broad set of reverse-engineering and security domains, each mapped to its own scenario skill: APK and iOS/mobile analysis, binary reverse engineering (exe/dll/so/elf) via IDA or radare2, .NET/C# decompilation, frontend JavaScript deobfuscation including custom VM/opcode reversal, malware analysis and YARA rule work, penetration testing and attack-chain orchestration, firmware and IoT targets, patch-diff N-day analysis, exploit development, EDR bypass research, API and GraphQL security, supply-chain and SBOM analysis, LLM/AI security, and CTF competitions through a dedicated 42-sub-skill sandbox orchestrator. Each of these is a real, distinct discipline with its own tooling and failure modes — the value of the routing layer is specifically in not treating them as interchangeable.
Built to Be Verifiable, Not Just Functional
The project backs its routing claims with actual regression testing rather than asking you to trust it: 41 routing rules validated against 163 bilingual benchmark cases, run in CI on both Windows and Ubuntu on every push. A separate structure-and-supply-chain coherence check specifically fails the build if an auto-install step isn't version-pinned — a real, concrete defense against a dependency silently changing under an agent mid-engagement. That level of testing discipline around routing logic is unusual for an Agent Skill and worth noting as a positive signal for a tool making decisions about which security tooling to invoke.
The most recent release, v1.0.1, is a good illustration of what that discipline looks like in practice rather than just as a stated policy. It moved the routing logic to a single structured configuration file (skills/config/routing.json) as the one source of truth instead of scattering routing decisions across client-specific instructions, grew the regression suite to the current 41-rule, 163-case bilingual benchmark, and added the read-only Case Review system described above. It also closed a specific consistency gap between the Bash and PowerShell versions of the authorization gate, so a case opened on Linux and reviewed on Windows — or vice versa — enforces the same scope rules either way. That's the kind of unglamorous cross-platform parity work that's easy to skip and expensive to get wrong in a tool whose entire job is gating access to a target.
Getting It Running
The setup is a plain git clone — there's no package to install or service to stand up first, which fits a tool meant to be dropped into an existing AI coding client's project directory. Three things need to already be present on the machine: a JDK (jadx and apktool depend on it), Node.js 22.12 or newer (the JS toolchain and MCP servers need it), and Python 3.x (Frida and the helper scripts run on it) — reverse-skill doesn't bundle any of these, it routes to them.
After cloning, the next step is running a platform-specific tool-index refresh script — PowerShell on Windows, Bash on Linux/macOS, and a separate variant for Kali specifically — which scans the machine for what's actually installed and writes the result to skills/tool-index.md. That index is what the routing layer checks before it recommends a tool: it's not assuming jadx or Frida is present, it's confirming it. Kali gets its own documented path since so much of the standard RE and pentest toolchain ships with it already, alongside separate platform docs for Ubuntu/Debian and macOS for anyone bootstrapping tools from scratch.
Client-Neutral by Design
The routing core, regression suite, and case workflow are deliberately kept independent of any specific AI coding client — Claude Code, Codex, Cursor, and OpenCode are all supported through their own adapter mechanisms rather than the routing logic being coupled to one tool's conventions. That mirrors the same portability principle behind well-built Agent Skills generally: one source of methodology, usable across whichever harness a team actually runs.
What to Weigh Before Using It
- This is unambiguously a professional/authorized-use tool, and should only be pointed at systems and applications you're authorized to test. The project's own scope-gating design assumes that authorization already exists before any action against a target — it's an orchestration layer for legitimate security work, not a substitute for having the engagement authorization in the first place.
- It assumes real tooling knowledge underneath. The routing layer picks the right methodology and tool, but interpreting IDA output, tuning a Frida hook, or judging whether a finding is actually exploitable still requires the underlying reverse-engineering or pentesting expertise — this speeds up an analyst, it doesn't replace one.
- License terms vary by submodule. The core project is MIT, but the bundled CTF Sandbox Orchestrator is GPLv3 and a referenced "Pentest Swarm AI" component is AGPL-3.0 and invoked externally rather than vendored — worth checking the specific license terms for any submodule before redistributing or building on top of this.
- It's a young project (created 2026) with active but still-forming community infrastructure — real CI and regression testing are a good sign, but expect the routing matrix and skill set to keep expanding rather than treating the current 41 rules as a finished, stable surface.
Practical Takeaway
reverse-skill is a well-structured answer to a real coordination problem in AI-assisted security work: an agent that's good at following a methodology still needs to be told which methodology applies, and reverse-skill's scope-gated, evidence-chained routing does that with more engineering discipline than a loose collection of prompts would. For security teams and researchers already running AI coding agents on authorized engagements or CTF work, it's worth evaluating specifically for the routing and audit-trail architecture — the parts that are hard to get right and easy to skip when building this kind of tooling in-house.
Teams building or securing AI-agent-driven workflows — including authorization gating, audit trails, and AI security architecture — can get hands-on help from Woyce Technologies.
FAQ
What is reverse-skill?
reverse-skill is an open-source Agent Skill that routes AI coding agents to the correct methodology and tooling for reverse engineering, authorized penetration testing, security research, and CTF tasks, with scope authorization gating and an evidence-chain audit trail built into its routing logic.
Is reverse-skill only for professional security researchers?
Its scope-gated design assumes authorization for any target action exists before the routing pipeline proceeds, making it built for authorized penetration testing, security research, and CTF contexts specifically — not a general-purpose or unrestricted hacking tool.
Which AI coding tools does reverse-skill work with?
It's designed to be client-neutral — Claude Code, Codex, Cursor, OpenCode, and other compatible AI coding clients can load it through their own adapter mechanisms, with the routing core kept independent of any one client's conventions.
What does the "scope gate" in reverse-skill actually do?
It requires a case to be initialized with confirmed authorization and a network profile before any action is taken against a target, structurally preventing the routing pipeline from acting on unauthorized scope.
Is reverse-skill free to use?
The core project is MIT-licensed. Some bundled or referenced components carry different licenses — the CTF Sandbox Orchestrator submodule is GPLv3, and a referenced Pentest Swarm AI component is AGPL-3.0 and invoked externally rather than included in the repo.
What kinds of security tasks does reverse-skill's routing cover?
A broad range, including APK and mobile reverse engineering, binary and firmware analysis, JavaScript deobfuscation, malware analysis, penetration testing and attack-chain work, exploit development, API and supply-chain security, LLM security, and CTF competition support.
What do I need installed before using reverse-skill?
A JDK for jadx and apktool, Node.js 22.12 or newer for the JS toolchain and MCP servers, Python 3.x for Frida and helper scripts, and a compatible AI coding client. reverse-skill routes to these tools rather than bundling them, so it checks for what's already on the machine via a per-platform tool-index refresh script.
How do I know which security tools reverse-skill has actually detected on my system?
Running the platform-specific refresh-tool-index script (PowerShell, Bash, or the Kali variant) scans the machine and writes the result to skills/tool-index.md, which the routing layer checks before recommending a tool for a given task.
What changed in reverse-skill's latest release?
Version 1.0.1 moved routing to a single client-neutral configuration file as the source of truth, expanded the regression benchmark to 41 rules and 163 bilingual test cases, added the read-only Case Review evidence system, and closed a consistency gap between the Bash and PowerShell versions of the authorization gate.
Where can I get help or report an issue with reverse-skill?
The project maintains a Discord community, a QQ group for Chinese-speaking users, and standard GitHub Issues for bug reports and feature requests.