⚒️ MCP Forge

Turn your browser clicks into AI superpowers instantly.

📖 The "Grandma" Explanation

Imagine you have an old recipe box. Every time you want a recipe, you have to manually flip through the cards.

Now, imagine you show a smart assistant exactly how you do it just once. The assistant watches you, takes notes, and says, "Got it!" From then on, you just say, "Hey, fetch me the apple pie recipe," and the assistant zips over, finds the card, and hands it to you.

MCP Forge is that assistant for your computer. You record yourself using a website (searching, clicking, following a workflow), hand that recording to MCP Forge, and it turns it into a permanent, reusable tool that an AI can use for you forever.

🎯 What is this project?

MCP Forge is an automated pipeline that ingests web traffic recordings (HAR files), analyzes the underlying API calls, and uses Large Language Models (Gemini 2.5 Flash / Claude) to instantly generate fully functional Model Context Protocol (MCP) servers.

These servers allow AI assistants (like Claude Desktop) to interact with web services autonomously, essentially giving your AI "hands" to use any website you've visited.

✨ Features

• 🔒 Security First: Automated regex redaction of Cookies, Authorization headers, and Bearer tokens to ensure your secrets never reach the LLM.
• 📡 Live Forging: A Server-Sent Events (SSE) streaming terminal in the UI that shows exactly what the "Smithy" is doing in real-time.
• ✅ AST Validation: Every generated Python script is parsed into an Abstract Syntax Tree before being saved, ensuring 100% syntactical correctness.
• 📦 One-Click Deployment: Download a complete .zip package containing your server, instructions, and the Claude Desktop configuration.
• 🎨 Cyber-Smithy UI: A high-performance, dark-mode dashboard built with Vite, React, and Tailwind CSS.

🚀 How it Works

1. Capture: Record your browser network traffic as a .har file using Developer Tools.
2. Intake: Upload the HAR and describe the task to the Cyber-Smithy Dashboard.
3. Scrubber: The backend filters noise (images, trackers) and redacts sensitive credentials.
4. Forge: Gemini/Claude analyzes the traffic and writes a clean, typed Python script using fastmcp.
5. Armory: View the code, copy the config, and download your ready-to-run AI tool.

Architecture Diagram

graph TD
    User([👤 User]) -->|Records Browser Network| HAR[(📄 HAR File)]
    HAR --> Frontend[⚛️ React Frontend]
    Frontend -->|Uploads HAR| Backend[⚡ FastAPI Backend]
    
    Backend -->|Redacts & Filters| Scrubber[🛡️ Security Scrubber]
    Scrubber -->|Filtered Endpoints| LLM((🤖 Claude / Gemini LLM))
    
    LLM -->|Generates Python Server| FileSys[(📁 Registry / Filesystem)]
    
    FileSys -->|Retrieves Server Details| Frontend
    
    FileSys -->|Server.zip| Claude[🧠 Claude Desktop]
    Claude -->|Calls MCP Tool| ExtAPI[🌐 External API]

🛠️ Tech Stack

• Backend: FastAPI, Python 3.14, uv
• AI engine: Google Gemini GenAI SDK
• Frontend: React, Vite, Tailwind CSS, Lucide Icons
• Output: Model Context Protocol (MCP) via fastmcp

💻 Getting Started

Prerequisites

• Python >= 3.14
• Node.js & npm
• uv package manager

1. Setup Backend

cd backend
uv sync
# Create .env and add: GEMINI_API_KEY="your_key"
uv run uvicorn main:app --reload

2. Setup Frontend

cd frontend
npm install
npm run dev

Visit http://localhost:5173 to start forging!

🛣️ Roadmap

• [x] Phase 1: Core Backend & HAR Parsing
• [x] Phase 2: AI Generation & Registry
• [x] Phase 3: Cyber-Smithy React Dashboard
• [ ] Phase 4: Autonomous Browser Agent (Direct URL recording)
• [ ] Phase 5: Multi-SDK Support (OpenCode, LangChain)