Supply Chain Digital Twin MCP

A Model Context Protocol server that constructs a live digital twin of multi-tier supply chain networks. Fuses corporate registry, trade flow, natural hazard, geospatial, and intelligence data from 17 Apify actors into a unified graph, then applies network science, game theory, and causal inference to simulate disruptions, optimize logistics, and forecast demand.

Tools (8)

#	Tool	Method	Price
1	simulate_disruption_cascade	Interdependent network percolation (Buldyrev et al. 2010) + cross-entropy importance sampling	$0.040
2	optimize_logistics_transport	Semi-discrete optimal transport, Kantorovich dual, Wasserstein distance	$0.040
3	model_adversarial_interdiction	Tri-level Stackelberg attacker-defender-attacker via Benders decomposition	$0.045
4	estimate_supplier_survival	Competing risks Cox proportional hazard model with cause-specific cumulative incidence	$0.035
5	reinforce_network_resilience	Algebraic connectivity (Fiedler λ₂) maximization via SDP relaxation	$0.040
6	compute_input_output_impact	Leontief inverse (I-A)⁻¹ via Neumann series, Ghosh supply-side multiplier	$0.035
7	identify_causal_disruption_paths	Fuzzy RDD with Imbens-Kalyanaraman optimal bandwidth selection	$0.035
8	forecast_multi_scale_demand	Rao-Blackwellized particle filter for hierarchical state-space decomposition	$0.030

Data Sources (17 actors across 5 categories)

Corporate (3): OpenCorporates, UK Companies House, GLEIF LEI Trade (3): UN COMTRADE, World Bank Indicators, Exchange Rates Hazard (5): USGS Earthquake, NOAA Weather, GDACS Disasters, FEMA, OpenAQ Air Quality Spatial (1): Nominatim Geocoder Intelligence (5): OFAC Sanctions, OpenSanctions, Censys, Website Changes, SAM.gov Contracts

Mathematical Foundations

Tool 1: Disruption Cascade — Interdependent Network Percolation

Models supply chains as interdependent networks where failure in one layer triggers cascades across layers. The percolation threshold p_c marks the critical fraction of node failures that triggers a first-order phase transition (catastrophic network collapse). Cross-entropy importance sampling efficiently estimates rare cascade probabilities by tilting the failure distribution toward high-impact events.

Tool 2: Logistics Transport — Semi-Discrete Optimal Transport

Solves the Kantorovich relaxation of the Monge problem: find the transport plan π that minimizes total cost ∫c(x,y)dπ(x,y) subject to marginal constraints. The dual formulation yields shadow prices (Lagrange multipliers) for each capacity constraint. The Wasserstein distance W₁(μ,ν) gives the minimum total transport cost between supply and demand distributions.

Tool 3: Adversarial Interdiction — Tri-Level Stackelberg

Formulates supply chain protection as a three-player sequential game: defender fortifies edges → attacker interdicts unprotected edges → operator routes flow through surviving network. Benders decomposition iteratively tightens upper and lower bounds until convergence to the Stackelberg equilibrium.

Tool 4: Supplier Survival — Competing Risks Cox Model

Models supplier failure as a competing risks process where multiple failure modes (financial distress, operational disruption, geopolitical instability) compete to cause exit. Cause-specific hazard functions h_k(t|X) = h₀_k(t)·exp(β_k'X) allow different covariates to drive different failure modes. The cumulative incidence function F_k(t) properly accounts for the probability of failing from cause k before being censored by other causes.

Tool 5: Network Resilience — Algebraic Connectivity

The Fiedler value λ₂ (second-smallest eigenvalue of the graph Laplacian L = D - A) measures how well-connected a network is. Higher λ₂ means faster mixing, better fault tolerance, and harder to disconnect. The SDP relaxation max λ₂(L + Σ w_ij·L_ij) subject to Σ w_ij ≤ budget finds the optimal set of edges to add for maximum resilience improvement.

Tool 6: Input-Output Impact — Leontief Inverse

The Leontief demand-driven model x = (I-A)⁻¹·d gives total output x required to satisfy final demand d, where A is the matrix of technical coefficients (a_ij = intermediate input from sector i per unit output of sector j). The Neumann series (I-A)⁻¹ = I + A + A² + ... converges when the spectral radius ρ(A) < 1. Forward linkages measure a sector's importance as a supplier; backward linkages measure its importance as a buyer. The Ghosh supply-side multiplier B = (I-G)⁻¹ captures output allocation effects.

Tool 7: Causal Disruption Paths — Fuzzy RDD

Regression discontinuity design exploits sharp thresholds in treatment assignment to estimate causal effects. The fuzzy variant handles probabilistic treatment: nodes near the disruption boundary receive treatment with probability between 0 and 1. The Imbens-Kalyanaraman bandwidth h* minimizes integrated MSE of the local linear estimator, balancing bias against variance. Identified confounders are separated from genuine causal propagation channels.

Tool 8: Multi-Scale Demand — Rao-Blackwellized Particle Filter

Decomposes demand into a hierarchical state-space model with trend, seasonal, and noise components at multiple time scales (daily → weekly → monthly → yearly). The Rao-Blackwellization marginalizes the linear-Gaussian substructure analytically (Kalman filter), using particles only for the nonlinear/non-Gaussian components. This dramatically reduces variance compared to a standard particle filter. The optimal base stock level is computed from the posterior predictive distribution to achieve the target service level.

Architecture

Client ──► Express (port 3018) ──► McpServer factory
                                       │
                          ┌─────────────┼─────────────┐
                          ▼             ▼             ▼
                    SSE transport  Streamable HTTP  Health endpoint
                          │             │
                          ▼             ▼
                   getOrBuildNetwork() (5-min TTL cache)
                          │
              ┌───────────┼───────────┐
              ▼           ▼           ▼
        17 Apify actors   Graph build   8 analysis tools
        (parallel fetch)  (fusion)      (per-request)

Usage

npm install
npm run build
npm start

Environment variables:

• APIFY_TOKEN — Apify API token for actor execution
• PORT — Server port (default: 3018)

Endpoints

• GET /sse — SSE transport for MCP
• POST /messages?sessionId=... — SSE message handler
• POST /mcp — Streamable HTTP transport
• GET /health — Health check with network stats