An AI Agent that analyzes your supply chain. 74 specialized tools. 8 SOP skills. Full audit trail.
-blue){kind=icon}
Ask questions. Agent analyzes data across inventory, cost, supplier, logistics, and risk. Every decision is traced and replayable.
You type: "库存健康检查" (check inventory health). The Agent:
- Classifies your intent (semantic router)
- Loads the inventory SOP skill
- Calls the right MCP tools (query_inventory → classify_abc_xyz → calculate_eoq)
- Synthesizes a structured report with [source][confidence] tags
- Persists the entire decision trace (auditable, replayable)
Not a dashboard. Not a chatbot. An Agent that does supply chain analysis.
# 1. Clone
git clone https://github.com/JiuTian-dev/SupplyChainCortex.git
cd SupplyChainCortex
# 2. Start
./start.sh
# 3. Open http://localhost:3000
# 4. Set DEEPSEEK_API_KEY in settings (gear icon)
# 5. Ask: "帮我做库存健康检查"| Capability | Why It Matters |
|---|---|
| 74 MCP Tools | Not just CRUD — EOQ with quantity discounts, Wagner-Whitin DP, Monte Carlo inventory simulation, multi-echelon safety stock, cascade risk propagation |
| FSM v2 Agent Engine | 6-state explicit state machine (classify→plan→execute→observe→decide→synthesize). No free-form LLM loops that drift |
| Progressive Tool Loading | 74 tools but only context-relevant ones sent to LLM. 60-80% token reduction |
| Knowledge Injection | Graph RAG (supply chain knowledge graph) + Agent Skills (8 SOP markdown files) loaded on trigger |
| Provider Adapters | DeepSeek V4 Flash, OpenAI GPT-4o, Anthropic Claude Sonnet 4.6, local Ollama — hot-swap one line |
| Full Audit Trail | Every decision traced: classify→plan→tools→synthesize. Replay from any node. EU AI Act + China AI compliance checklist |
| 3-View Layout | Chat-first interface with slide-out data panel. Legacy dashboard panels preserved for quick reference |
User Query
↓
Semantic Router (LLM-based intent classification)
↓
FSM v2 Agent Engine (6-state explicit state machine)
↓
Provider Adapter Layer (DeepSeek / OpenAI / Anthropic / Ollama)
↓
Progressive Tool Filter (74 tools → intent-relevant subset)
↓
Knowledge Injection (Graph RAG + 8 SOP Skills)
↓
MCP Tool Execution (parallel, with autonomy policy)
↓
Trace Persistence (DecisionTrace → TraceStep → TraceToolCall + TracedClaim)
↓
Synthesize + Done (MARC-tagged, source-labeled, audit-traced)
| Domain | MCP Tools | What Agent Can Do |
|---|---|---|
| Inventory | 20 tools | ABC-XYZ classification, EOQ with discounts, safety stock, Monte Carlo simulation, DRP, Wagner-Whitin, newsvendor model, fill rate analysis, warehouse capacity |
| Cost | 11 tools | Landed cost breakdown, commodity price tracking (copper/aluminum/steel/plastics), SCFIS freight rates, carbon pricing, exchange rates, break-even analysis, optimal pricing |
| Supplier | 8 tools | Performance scoring, 6-month trend analysis, geographic distribution, cascade risk from supplier failure, 1688/Alibaba discovery |
| Logistics | 5 tools | Shipment tracking, port congestion monitoring, marine weather, transport route optimization, status updates |
| Risk | 5 tools | Cascade risk propagation (8 scenarios), CPSC recall monitoring, coherence audit, weather risk, recall risk prediction |
| Market | 8 tools | Amazon competitor analysis, brand sentiment, cross-platform arbitrage, product feed generation |
| Operations | 7 tools | Reorder creation (batch), stock transfer, inventory adjustment, supplier CRUD, shipment status workflow |
| Universal | 10 tools | Dashboard metrics, analytics, decision graph, financial simulation, sandbox simulation, workflow execution, web search, chart generation |
Progressive-disclosure markdown skills that encode supply chain expertise. Triggered when the user's query matches:
| Skill | Trigger | What It Does |
|---|---|---|
inventory-health-check |
"库存健康" | Overview → ABC → slow-moving → risk → KPIs → recommendations |
cost-optimization |
"成本优化" | Commodities → FX → freight → tariff → landed cost → break-even |
supplier-risk-assessment |
"供应商风险" | Performance → trends → geography → cascade → scoring |
logistics-port-monitor |
"物流" "货运" | Stats → shipments → port congestion → weather → ETA risks |
tariff-trade-war-sim |
"关税" "贸易战" | Current tariffs → simulate → carbon price → financial impact → sandbox |
full-health-report |
"全健康报告" | Dashboard → analytics → risk → cascade → charts |
procurement-planning |
"采购计划" | Reorder list → EOQ → DRP → batch reorder |
compliance-audit |
"合规" "认证" | Compliance check → CPSC → recall risk → carbon → coherence audit |
| Layer | Technology |
|---|---|
| Framework | Next.js 16 (App Router, Turbopack) |
| Language | TypeScript 5 (strict) |
| Database | PostgreSQL 16 + Prisma ORM (28 models) |
| Agent Engine | FSM v2 (6-state, model-agnostic) |
| Math Engines | Python 3 + NumPy (24 OR models) via bridge |
| Search | SearXNG self-hosted (70+ engines) |
| Streaming | SSE (Server-Sent Events) |
| UI | Tailwind CSS 4 + shadcn/ui + Recharts |
| Testing | Vitest (583 tests) + Playwright |
| Skills | 8 SKILL.md files (progressive disclosure) |
| Cache | ICacheBackend (Memory + PostgreSQL backends) |
- Node.js 20+ / Bun 1.0+
- PostgreSQL 16 (Docker:
docker compose up -d postgres) - Python 3.8+ (for math engines)
- DeepSeek API Key (or OpenAI / Anthropic / local Ollama)
bun install
bun run dev # http://localhost:3000
bun run test # 583 tests
npx tsc --noEmit # zero new errors
bun run db:push # sync Prisma schema
bun run db:seed # seed 50 products + 11K sales recordsCC BY-NC 4.0 — Free for non-commercial use. Commercial licensing available.