What they built

Typus MCP: The Enterprise Compliance Gateway for Agentic Banking

FinGate MCP: Enterprise Compliance Gateway With EU AI Act enforcement looming, financial institutions face an infrastructure crisis. Standard MCP servers act as unmonitored open pipes between AI and banking APIs, lacking the human oversight mandated for High-Risk Systems. FinGate MCP is a proxy layer between orchestration clients and downstream tools, enforcing strict structural validation, security, and human mediation. What It Does & Who It Is For It transforms AI ecosystems into compliant architectures by intercepting high-risk operations, forcing explicit authorization, and logging actions. Built for BFSI Compliance Officers ensuring regulatory adherence, Security Engineers blocking prompt injections, and AI developers deploying workflows. Four Security Pillars 🛡️ HITL Mediation: High-risk operations (e.g., $10K transfers) trigger an interception gate. The system suspends execution and routes an approval prompt via /api/mediate before edge-native rehydration. 🧠 Threshold Tuning: Using Bayesian Optimization, FinGate dynamically tunes risk thresholds based on historical approval patterns, reducing false positives while keeping strict 95th-percentile safety limits. 🔒 Cryptographic Verification: To block Man-in-the-Middle exploits, FinGate uses SHA-256 intent hashing. Human-approved parameters are cryptographically mapped directly to the backend payload. 📜 Audit Logging: Every intervention and user action is immutably logged via a SQLite Prisma model (MediationLog) for non-repudiable regulatory compliance. Domain-Driven Architecture Host App (Next.js): Handles stream routing (src/). Domains: agent/ (AI logic), security/ (HITL/Bayesian engine), infrastructure/ (Prisma bounds), and interfaces/ (TS definitions). MCP Server (NitroStack): The secured endpoint (mcp-server/) containing banking tools and strict Zod validations. Run bun run scripts/pentest.ts to test adversarial payloads and cryptographically prove intercept compliance.