GlobalFUN — Cross-Border E-Commerce Multi-Agent System

A LangGraph + MCP Protocol powered multi-agent system for cross-border e-commerce operations.

Overview

GlobalFUN is an AI-powered platform that automates the full lifecycle of cross-border e-commerce — from demand analysis and market research to product selection, content generation, compliance checking, and operations strategy. Unlike traditional chatbots, it uses a 6-agent collaboration architecture with structured card-based output.

Architecture

User Query
    │
    ▼
┌─────────────┐
│ Intent Parser│  ← Intent classification + Entity extraction
└──────┬──────┘
       │  Conditional Routing
       ▼
┌──────────────────────────────────────┐
│  ┌──────────┐ ┌──────────┐ ┌───────┐ │
│  │ Market   │ │ Product  │ │Compli-│ │  ← Parallel execution
│  │ Analyst  │ │ Scout    │ │ance   │ │
│  └────┬─────┘ └────┬─────┘ └───┬───┘ │
│       └─────────────┼───────────┘     │
│                     ▼                 │
│  ┌──────────┐ ┌──────────┐           │
│  │Copywriter│ │ Strategy │           │  ← Content & Strategy
│  └────┬─────┘ └────┬─────┘           │
│       └─────────────┼───────────┘     │
│                     ▼                 │
│              ┌────────────┐           │
│              │  Compiler  │           │  ← Structured output
│              └────────────┘           │
└──────────────────────────────────────┘

Agents

Agent	Role	Icon	Tools Used
Intent Parser	Intent Recognition	🔍	LLM + keyword fallback
Market Analyst	Market Research	📊	product_search, currency
Product Scout	Product Selection	🛍️	price_compare, review_analyze
Compliance	Regulatory Check	🛡️	compliance_check
Copywriter	Content Generation	✍️	LLM generation
Strategy	Operations Strategy	📈	LLM + market data

MCP Tools

Tool	Description
price_compare	Multi-platform price comparison
product_search	Product search & discovery
logistics	Logistics & tariff calculation
compliance_check	Regulatory compliance verification
currency	Real-time currency conversion
review_analyze	Product review analysis

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Set up environment
cp .env.example .env
# Edit .env with your API key

# 3. Run CLI demo
python examples/run_demo.py
python examples/run_demo.py "Compare SK-II prices across platforms"

# 4. Run tests
python -m pytest tests/ -v

Project Structure

cross-border-ai-agent/
├── src/
│   ├── config.py              # Global settings
│   ├── agents/                # 6 specialized agents
│   │   ├── intent_parser.py   # Intent classification
│   │   ├── market_analyst.py  # Market research
│   │   ├── product_scout.py   # Product evaluation
│   │   ├── copywriter.py      # Content generation
│   │   ├── compliance.py      # Regulatory checks
│   │   └── strategy.py        # Operations strategy
│   ├── tools/                 # MCP-compatible tools
│   │   ├── price_compare.py
│   │   ├── product_search.py
│   │   ├── logistics.py
│   │   ├── compliance_check.py
│   │   ├── currency.py
│   │   └── review_analyze.py
│   ├── graph/                 # LangGraph workflow
│   │   ├── state.py           # Shared state definition
│   │   └── workflow.py        # Graph construction
│   └── mcp/                   # MCP protocol layer
│       ├── protocol.py        # Message types
│       └── router.py          # Tool routing
├── tests/
│   └── test_pipeline.py
├── examples/
│   └── run_demo.py            # CLI demo
└── demo/
    └── work/
        └── cross-border-ai.html  # Interactive web demo

Key Design Decisions

1. LangGraph over LangChain chains — StateGraph enables conditional routing and parallel execution, which is critical for multi-agent systems.

2. MCP Protocol — Tools are decoupled from agents via a protocol layer. Any agent can call any tool through the router, making the system extensible.

3. Keyword fallback — Intent parsing has a keyword-based fallback when LLM is unavailable, ensuring the system works offline.

4. Structured state — PipelineState TypedDict provides type safety and clear data flow between agents.

5. Structured card output — Results are delivered as structured cards (price comparison, compliance report, copywriting, strategy) rather than plain text.

Tech Stack

• LangGraph — Multi-agent orchestration
• MCP Protocol — Tool abstraction layer
• Python 3.11+ — Runtime
• FastAPI — API server
• React + TypeScript — Frontend UI

License

MIT