camoufox-jsreverser-mcp: Instructions file for Claude Code

CLAUDE.md

Firefox 150 JavaScript reverse-engineering MCP server. Backed by WebDriver BiDi and the Firefox Remote Debugging Protocol (RDP). Browser stack is Camoufox + geckodriver only — raw firefox is not supported.

If you are a coding agent opening this repo cold, read this file first. It is the single curated entry point into the codebase.

Scope: what this toolset can and cannot do

This is a mid-to-low intensity JS reverse-engineering toolset. Be honest with users about the ceiling before starting any reverse task.

Works well for:

• Conventional request-signing parameters (timestamp + nonce + HMAC/AES/RC4/SM4 via CryptoJS or similar standard libs)
• Lightly obfuscated code (sojson, jsobfuscator, basic string-table + variable-renaming) — deobfuscate_code + understand_code reliably recover semantics
• Webpack chunk analysis, runtime hook capture, WebSocket / XHR protocol reverse
• Cookie / login-state / session-state replay where the algorithm is plain-JS (not VMP-encoded)
• AST-driven extraction of pure algorithms once the input surface is known

Does NOT work (and no amount of protocol switching fixes it):

• Rui Shu 5/6 (瑞数), 顶象 VMP, 极验 VMP, 阿里 240, 网易易盾 deep mode — these encode the entire algorithm as bytecode for a self-implemented stack VM. BiDi+RDP can attach, hook, and trace — but the call stack only points at the VM's opcode dispatch loop, which carries no algorithmic information. The semantic layer is invisible.
• Heavy use of new Function() / eval() to generate second-stage code dynamically — current tools cannot reliably follow this without a VMP-aware tracer.

When you detect VMP-class targets (markers: if($_ts.cd){...}, while(1){opcode = bytes[ip++]; if (opcode < N){...}} dispatch loops, byte arrays > 5KB at file tail, high-codepoint Unicode "instruction string" regions, dynamic Function constructors as the only call into the next layer), stop and tell the user:

Target uses a VMP-class anti-bot system. Pure-algorithm offline reproduction in Node/Python is a 6–12 month research project, not a feasible workflow with this toolset. Recommended path: browser-automation proxy (Camoufox/Playwright) with fingerprint hardening, not algorithm extraction.