Deobfuscate MCP Server

An LLM-Optimized Model Context Protocol (MCP) server designed to help Large Language Models reverse-engineer, navigate, and understand minified and bundled JavaScript code.

Why this Server?

Standard "beautifiers" only format code, leaving LLMs to struggle with massive, linear text files that overflow context windows. This server treats minified code as a searchable database, allowing LLMs to:

1. De-obfuscate & Unbundle: Uses webcrack to split Webpack/Browserify bundles into individual modules.

2. See the Architecture: Returns a JSON summary of the file's structure (exports, functions) before reading the code.

3. Navigate Randomly: Read specific modules on-demand (get_module) instead of the whole file.

4. Search Semantically: Find where specific strings or patterns exist across thousands of unbundled modules.

Features

• deobfuscate: The entry point. Unpacks bundles and caches them in memory.

  • returnCode (default false): Returns a summary by default to save context. Set to true to see the full code.

  • skipVendor: Filter out node_modules and webpack boilerplate to focus on app code.

  • mangle: Shorten variable names to save tokens.

  • jsx: Restore React JSX syntax.

• analyze_structure: Returns a high-level AST summary (functions, classes, exports) to save tokens.

• list_modules: Lists all modules found in the unpacked bundle.

• list_functions: Scans cached modules to list defined functions/classes with signatures and parameters.

• get_call_graph: Generates a call graph for a specific function, identifying what it calls and what calls it.

• get_module: Fetches the formatted source code of a single module.

• get_symbol_source: Extracts only a specific function, class, or variable to save tokens.

• search_modules: Regex/String search across all modules.

• format_code: Standard Prettier formatting for JS/HTML/CSS.

• get_help: Returns detailed documentation for any tool.

Installation & Setup

Install from npm (Recommended)

Build from source

Pack and Install (Optional)

To create a distributable tarball and install it globally:

Client Configuration

To use this server with your favorite LLM client, add the following configuration.

Claude Code

Run the following command in your terminal:

Alternatively, edit ~/Library/Application Support/Claude/claude_desktop_config.json:

Gemini CLI

Run the following command in your terminal:

Alternatively, edit ~/.gemini/settings.json:

Antigravity

Run the following command in your terminal:

Alternatively, edit ~/.antigravity/settings.json:

Development / Local Usage

If you are running the server from the source code, replace the command with:

(Replace

Usage

Start the server:

Example Workflow for an LLM

1. User: "Analyze this minified file: bundle.min.js..."

2. LLM: Calls deobfuscate(code="...", skipVendor=true).

   • Server: "Deobfuscation complete. Unbundled 15 modules. Skipped 45 vendor modules..."

3. LLM: "Okay, I see the entry point requires module 42. What is that?"

4. LLM: Calls get_module(id="42").

   • Server: Returns code for module 42.

5. LLM: "Where is the 'login' function defined?"

6. LLM: Calls search_modules(query="function login").

   • Server: "Found in module 88."

7. LLM: "I just want to see the login function logic."

8. LLM: Calls get_symbol_source(symbolName="login", moduleId="88").

   • Server: Returns only the source for the 'login' function.

Development

• Build: npm run build

• Test: npm test

Limits

• File Size: The server accepts input files (bundles) up to 50MB.

• Memory: Unbundled modules are cached in RAM. Very large bundles (hundreds of MBs unpacked) may exhaust the server's available memory depending on your environment.