reverse-skill: The AI Agent Skill Router for Reverse Engineering and Security Research (2026 Guide)
reverse-skill is a free, open source cybersecurity skill router for AI agents. 41 routing rules, 163 benchmark cases, supports APK, binary, JS, CTF, pentesting. Works with Claude Code, Cursor, Cline. 22K+ stars.
AI agents are bad at security research. When an agent encounters an APK, a binary, encrypted frontend JavaScript, or a CTF challenge, it guesses commands. It does not know whether to use jadx, apktool, Frida, IDA, or Burp Suite for a given task. APK, ELF, JS, PCAP, and CTF tasks each need different playbooks, and tools, MCP servers, and scripts are scattered across machines. The same mistakes get repeated because experience is not reused.
reverse-skill, developed by zhaoxuya520, is a free, open source cybersecurity skills router for AI coding agents. When an AI agent (Claude Code, Codex, Cursor, Cline, or another compatible client) encounters a security research task, this package routes it to the right methodology, checks available tools, and executes a repeatable workflow instead of guessing. 41 routing rules across 163 benchmark cases, 42 tracked modules, cross-platform CI validation. 22,400+ stars on GitHub.
In this guide, you'll learn what reverse-skill is, who it's for, and how to install it for your AI agent.
What is reverse-skill?
reverse-skill is a cybersecurity skills router that gives AI agents structured playbooks for reverse engineering, penetration testing, and security research. Instead of letting the agent guess which tool to use, reverse-skill routes the task through a structured pipeline:
- The agent reads
RULES.md(global routing rules with a scope gate before any action) MASTER-ROUTING.mdperforms primary triage (readsrouting.jsonwith 41 rules, R0 through R40)case-initcreates a case directory with scope, timeline, and work items- The matched scenario skill loads the right tools, MCP servers, and scripts
- The agent executes the workflow, building a timeline with evidence, findings, and paths
- Results are compiled into a report and field journal
The routing core is driven by one structured configuration (routing.json), validated by cross-platform CI (Windows and Ubuntu), and kept separate from optional client adapters. This means the routing logic, regression suite, manifests, and case workflow do not depend on a specific AI client.
Who is it for?
- Security researchers and reverse engineers: Professionals who use AI agents to help with APK analysis, binary reversing, malware analysis, or CTF challenges and want the agent to follow proper methodology instead of guessing.
- Penetration testers: Ethical hackers who want their AI agent to follow a structured pentesting workflow with scope gates, evidence chains, and proper reporting.
- CTF competitors: Players who want an AI agent that can handle 42 different CTF sub-skills, from crypto to pwn to web challenges.
- AI agent developers: Builders creating security-focused agent workflows who want a pre-built, tested routing system with 163 benchmark cases.
What makes reverse-skill different from asking an AI agent to "analyze this binary"?
- 41 routing rules with 163 benchmark cases: Every routing decision is tested. The regression suite runs 163 cases (hint to expected primary route) and fails CI on any mismatch. This is not a prompt. It is a tested routing system.
- Scope gate before action: The agent cannot take action on a target until the scope is defined (authorization, network profile). This prevents accidental testing of out-of-scope systems.
- 42 tracked modules across 20 scenarios: APK reverse, iOS mobile, binary reverse (IDA, radare2), .NET, frontend JS, DSL VM, HTTP capture, malware analysis, pentest tools, attack chain, case review, CTF (42 sub-skills), firmware, patch diff, pwn, EDR bypass, API security, supply chain, LLM security, and more.
- Evidence chain and field journal: Every action is logged in a timeline with evidence, findings, and paths. The case review module provides read-only evidence graph review and artifact fixity checks. This is proper forensic methodology, not ad-hoc commands.
- Client-neutral: Works with Claude Code, Codex, Cursor, Cline, OpenCode, and any compatible client. The routing core does not depend on a specific AI client. Client-specific configuration is optional and outside the core routing contract.
- Self-evolving knowledge base: The system learns from completed cases. Experience is reused instead of repeating the same mistakes.
- Free and open source: MIT license (CTF sub-module is GPLv3, pentest swarm is AGPL-3.0 invoked via CLI). 22,400+ stars, 20 contributors.
What you need before you start
- An AI coding agent: Claude Code, Codex, Cursor, Cline, OpenCode, or another compatible client.
- Java / JDK: For jadx and apktool (Android analysis tools).
- Node.js 22.12+: For JS toolchain and MCP servers.
- Python 3.x: For Frida and helper scripts.
- Security tools: IDA Pro, radare2, Ghidra, Frida, Burp Suite (the router detects which are installed and routes accordingly).
Step-by-step installation
Step 1: Clone the repository
git clone https://github.com/zhaoxuya520/reverse-skill.git
Step 2: Refresh the tool index
The router needs to know which security tools are installed on your machine. Run the platform-specific refresh script:
Windows:
powershell -File skills/scripts/refresh-tool-index.ps1
Linux / macOS:
bash skills/scripts/refresh-tool-index.sh
Kali Linux:
bash kali/scripts/refresh-tool-index.sh
Check skills/tool-index.md to see which tools were detected.
Step 3: Load the skill in your AI agent
For Claude Code, the repository is loaded as a project skill. For Cursor, Cline, or other clients, follow the client's adapter or project-instruction mechanism. The routing core is client-neutral.
Step 4: Give your agent a task
Ask your agent to analyze a target. The router will triage the task, match it to the right scenario skill, check available tools, and execute a structured workflow:
Analyze this APK: ./target.apk
The agent will route through the APK reverse skill, use jadx and apktool if available, extract findings, and compile a report.
Common errors and how to fix them
| Error | What it means | How to fix it |
|---|---|---|
| Routing mismatch in CI | A routing rule change caused a benchmark case to route to the wrong skill. | Run test-routing.ps1 locally to identify the failing case, fix the routing rule in routing.json, and re-run. |
| Tool not detected | The tool index does not show a tool you have installed. | Manually run the refresh script for your platform. Ensure the tool is in your PATH. Check tool-index.md for detected tools. |
| Agent ignores routing rules | The AI client is not loading the repository as a skill or project instruction. | Check the client's documentation for how to load external skills. For Claude Code, ensure the repo is in the skills directory. |
| Scope gate blocks action | The agent is trying to act on a target before the scope is defined. | Complete the case-init step first (define authorization and network profile) before asking the agent to take action. |
reverse-skill vs asking an AI agent directly
| Feature | reverse-skill (structured router) | Asking agent directly |
|---|---|---|
| Tool selection | 41 rules route to the right tool automatically | Agent guesses, often wrong |
| Methodology | Structured workflow with scope gate, evidence chain, report | Ad-hoc commands, no methodology |
| Repeatability | 163 benchmark cases, CI-validated | Non-repeatable, different every time |
| Evidence | Timeline, findings, paths, field journal | No structured evidence |
| Scope safety | Scope gate prevents out-of-scope action | No safety mechanism |
| Experience reuse | Self-evolving knowledge base | Starts from scratch every time |
| Best for | Professional security research and CTF | Quick one-off questions |
Bottom line: reverse-skill gives AI agents proper cybersecurity methodology. Instead of guessing which tool to use and producing ad-hoc results, the agent follows a tested routing system with 41 rules, 163 benchmark cases, scope gates, and evidence chains. If you use AI agents for reverse engineering, penetration testing, or CTF competitions, this is the most comprehensive open source skill router available. The 22,400+ stars and active development community confirm it is solving a real problem that every security researcher using AI agents faces.
3 alternatives worth checking out
- Pentest Swarm AI: A related project that reverse-skill invokes via CLI or MCP. Pentest Swarm AI focuses on automated penetration testing orchestration. It is AGPL-3.0 licensed and reverse-skill calls it as an external tool rather than including its source.
- BurpSuite MCP (github.com/harishm111/burpsuite-mcp): An MCP server that exposes Burp Suite functionality to AI agents. While reverse-skill is a routing system, BurpSuite MCP is a single tool integration. Use them together: reverse-skill routes to the right scenario, and BurpSuite MCP provides the HTTP interception tool.
- Ghidra (github.com/NationalSecurityAgency/ghidra): The NSA's open source reverse engineering framework. Ghidra is a tool, not an agent skill. reverse-skill can route to Ghidra when it detects it is installed. Use Ghidra for the actual binary analysis and reverse-skill to tell your agent when and how to use it.
Found this guide useful? Check out more AI tools and open source projects on Sudo Scout.
Related posts
Prime Agent: The Self-Improving AI Coding Agent With Persistent IPython and Built-In Subagents (2026 Guide)
Prime Agent is a free, open source self-improving RLM agent for coding and long-running autonomous tasks. Persistent IPython, built-in subagents, continual harness refinement, daemon-backed sessions. 15.8K stars.
NVIDIA Switchyard: Route LLM Traffic Across Models and Providers Without Changing Your API (2026 Guide)
NVIDIA Switchyard is a free, open source Rust proxy that routes LLM requests across providers. Translates between OpenAI and Anthropic APIs, supports Claude Code and Codex, A/B benchmarking, cost optimization. Apache 2.0.
Needle 2: The 14MB AI Model That Runs on Phones, Wearables, and Smart Home Devices (2026 Guide)
Needle 2 is a free, open source 45M-parameter foundation model for tool calling on tiny devices. 14MB binary, 28MB RAM, runs fully offline. LoRA fine-tuning, confidence gating, structured extraction. 5.4K stars.