AI agents that can actually reverse-engineer JavaScript

What it does

js-reverse-mcp plugs into AI coding assistants (Claude, Cursor, Copilot, etc.) via the Model Context Protocol and gives them 21 tools to debug, breakpoint, and analyze JavaScript running in real web pages. It runs a visible Chrome instance with persistent login state, so your agent can step through minified code, inspect scope variables, trace XHR call stacks, and intercept WebSocket messages like a human reverse engineer would.

The interesting bit

The project treats anti-detection as a first-class concern, not an afterthought. It layers Patchright’s CDP-protocol evasion (avoiding Runtime.enable and other automation leaks) with an optional --cloak mode that swaps in a custom Chromium binary patched at the C++ level for 49 fingerprinting vectors — canvas, WebGL, audio, GPU, fonts, the works. The wrapper itself injects zero JavaScript, because Object.defineProperty hacks are themselves detection signals.

Key highlights

• 21 MCP tools covering navigation, script analysis, breakpoints, execution control, network tracing, and WebSocket inspection
• Dual-layer anti-detection: protocol-level (default) plus optional source-level binary patches for hardened targets like Cloudflare Turnstile or DataDome
• Persistent profiles for login state, with --isolated for clean sessions and --browserUrl to attach to existing Chrome or third-party fingerprint browsers
• “CDP silent navigation” defers activating Network.enable / Debugger.enable until a tool actually needs them, reducing early-page fingerprint exposure
• Rebuilt from chrome-devtools-mcp with agent-first tool design

Caveats

• --cloak mode downloads a ~200MB custom Chromium binary on first run; the README warns this can look like the MCP has frozen
• Connecting to third-party fingerprint browsers (AdsPower, BitBrowser) requires extracting a CDP endpoint via each vendor’s local API, with per-launch random ports
• The tool exposes full browser content to the MCP client; the authors explicitly warn against using it on pages with sensitive data

Verdict

Worth a look if you’re building AI agents that need to analyze or interact with production web apps that fight back against automation. Skip it if your use case is static code analysis or you don’t need to evade bot detection — a simpler headless setup will be less moving parts.