Skip to main content

​
1. Executive Summary

The SINT MCP Marketplace is an advanced platform for managing and executing Modular Cloud Programs (MCPs). It is designed for high-security applications where modular, on-demand cloud logic must be discoverable and executable by authenticated users or AI agents. The system integrates advanced security models, serverless execution, developer-friendly onboarding, and future Web3 and edge computing capabilities.

​
Key Capabilities:

  • Confidential Execution: Hardware-backed compute enclaves (AWS Lambda, KMS encryption, IAM isolation).
  • Zero Trust Security: Multi-layer authentication, granular token scoping, and audit logging.
  • Scalability: Auto-scaling serverless infrastructure with provisioned concurrency for low-latency use cases.
  • Extensible API-first Approach: OpenAPI specifications for each MCP.
  • Developer and AI Agent Support: Ready for autonomous agents and human developers.

​
2. System Architecture

​
2.1 Layered Design

  1. Presentation Layer
    • Web (Next.js) UI, CLI tools, chatbot interfaces.
    • Handles login, service connections, and MCP invocation.
  2. API Gateway Layer
    • AWS API Gateway routes and validates requests.
    • Uses Cognito/Keycloak JWT validation.
  3. Orchestration & Business Logic Layer
    • Node.js microservice managing routing, session tokens, scope validation, and MCP matching.
  4. Execution Layer
    • AWS Lambda, one per MCP for strict isolation.
    • Stateless runtime for ephemeral execution.
  5. Data Layer
    • PostgreSQL for metadata and relational data.
    • DynamoDB for fast lookups and caching.
    • AWS Secrets Manager for encrypted token storage.
  6. Monitoring & Logging Layer
    • CloudWatch (logs and metrics), Datadog (dashboards), X-Ray (distributed tracing).

​
2.2 Core Interaction Flow

  1. User logs in through Keycloak → obtains JWT session token.
  2. User connects a service (e.g., Google) → OAuth token stored in encrypted vault.
  3. User selects MCP → system verifies required scopes.
  4. MCP request routed through API Gateway → Lambda invoked.
  5. Lambda executes business logic using scoped tokens.
  6. Result returned, logged, and metrics updated.

​
3. Security Model

​
3.1 Authentication & Authorization

  • Identity Provider: Keycloak, configured for OAuth 2.1 PKCE.
  • Session Tokens: Short-lived JWT with refresh token rotation.
  • Third-Party Tokens: Separate encrypted storage per user/service.
  • Access Scopes: MCP-specific scope declarations enforced at runtime.

​
3.2 Execution Isolation

  • Each MCP Lambda has:
    • Dedicated IAM role with least privilege.
    • VPC-based egress filtering.
    • No persistent storage (ephemeral only).
  • AWS Shield + WAF to prevent DDoS attacks.

​
3.3 Data Protection

  • AES-256 envelope encryption for stored tokens and secrets.
  • All API traffic over HTTPS/TLS 1.3.
  • Structured logging with PII redaction.

​
4. Data Model & Storage

​
4.1 Entities

  • Users: Core identity and session data.
  • MCP Registry: Metadata, OpenAPI spec links, tags, and owner.
  • Authorizations: Linked third-party tokens with scope/expiry.
  • Executions: Invocation payloads, responses, and performance metrics.
  • Feedback: Ratings and developer-provided comments.
  • Logs: Structured event logs with request IDs.

​
5. DevOps Pipeline

​
5.1 CI/CD Steps

  1. Developer Commit → GitHub Actions triggered.
  2. Code Linting & Testing → Pre-build validation.
  3. OpenAPI Validation → Ensures MCP spec compliance.
  4. Security Scan → SAST, dependency vulnerability check.
  5. Build Lambda Package → Dockerized or ZIP build.
  6. Deploy to Staging → Blue/green strategy.
  7. Integration Tests → Automated runtime validation.
  8. Promote to Production → Canary or full release.

​
5.2 Observability & Incident Response

  • CloudWatch Alarms for latency and error spikes.
  • Datadog dashboards for live performance.
  • Slack PagerDuty integration for incidents.
  • Sentry for front-end and API error tracking.

​
6. Edge Cases & Failure Modes

  • Expired Tokens: Automatic refresh, fallback to reauthorization.
  • OAuth Consent Denial: MCP invocation blocked with user notification.
  • Cold Start Delays: Provisioned concurrency and scheduled warmups.
  • Registry Outage: Cache-based failover for MCP metadata.
  • Execution Timeout: Lambda configured for 30s max execution with fallback messaging.

​
7. API Specifications (Condensed)

​
Endpoints

  • POST /auth/login – Keycloak login.
  • GET /auth/callback – Token exchange.
  • GET /mcp – List MCPs.
  • GET /mcp/ – Get MCP metadata.
  • POST /mcp//invoke – Execute MCP.
  • GET /profile/services – View service connections.
  • POST /profile/services//connect – Link third-party service.
⌘I