Detailed Comparison: MCP Gateway Features

This document provides a technical deep dive into the Model Context Protocol (MCP) implementations in Light-Fabric and AgentGateway.

Feature Matrix

1. Architectural Intent

AgentGateway: The Network Proxy Layer

AgentGateway is designed as a high-availability proxy for MCP servers. Its primary focus is the North-South traffic between an application and multiple MCP backends.

• Multiplexing: Optimized for merging multiple MCP backends into a single upstream connection (mergestream.rs).
• Protocol Translation: Excels at translating between SSE, Streamable HTTP, and WebSocket transports.
• Infrastructure Focus: Operates as a Kubernetes-native component managed via manifests and standard networking policies.

Light-Fabric: The Managed Enterprise Platform

Light-Fabric provides a Unified Governance Fabric that treats AI agents and MCP tools as part of the broader enterprise API ecosystem.

• Unified Gateway: The AI Gateway (Rust/Pingora-based) serves as a single entry point for UI, Agents, and Tools, supporting both MCP and traditional REST/RPC APIs.
• Centralized Portal: Uses the Light-Portal as a control plane for onboarding (auto-discovery), configuration (hot-reloading), and security management.
• Governed Intelligence: Integrates the gateway directly with Hindsight Memory and the Fine-Grained Rule Engine, ensuring that every tool call is governed by corporate compliance rules (e.g., row/column masking).
• End-to-End Security: Maintains a single JWT-based identity from the user's chat interface all the way to the underlying MCP or API endpoint.

2. Security & Authorization

AgentGateway: Infrastructure-Aware RBAC

AgentGateway uses Common Expression Language (CEL) for its authorization policies.

• Capabilities: High-speed, network-level blocking based on JWT claims and request headers.
• Limitation: Lacks native support for content-aware data masking or organizational hierarchy logic.

Light-Fabric: Content-Aware Managed Auth

Light-Fabric provides a mature Fine-Grained Authorization layer:

• Managed ABAC/PBAC: Supports Role, Group, Corporate Position (Hierarchy), and Attribute-based protection.
• Data Privacy: Supports native Row and Column filtering (data masking), ensuring agents only see data they are authorized to process.
• End-to-End JWT: The same JWT token is propagated from the UI through the Agent to the AI Gateway and MCP tool.

3. Lifecycle & Tool Onboarding

AgentGateway: Configuration-Driven

Onboarding tools in AgentGateway is an infrastructure task:

• Manual Mapping: Requires defining Kubernetes Custom Resources (HTTPRoute, Backend) to map MCP servers to the gateway.
• Scope: Primarily focused on exposing existing MCP servers.

Light-Fabric: Registry-Driven

Light-Fabric provides a "Zero-Effort" onboarding experience via Light-Portal:

• Auto-Discovery: Registering an MCP API triggers an automatic tools/list call to populate the registry.
• Protocol Transformation: Automatically transforms existing OpenAPI/REST and RPC services into MCP tools without requiring wrappers.
• Centralized Governance: All tools (Native, REST, MCP) are managed in a single unified registry.

4. Control Plane & Configuration

AgentGateway: Kubernetes-Native

• Orchestration: Managed via the Istio/xDS control plane.
• Updates: Configuration changes are applied via Kubernetes manifests (YAML).

Light-Fabric: Portal-Managed

• Hot-Reloading: Uses a dedicated Config Server and Control Plane to update gateway and agent configurations in real-time without restarts.
• Enterprise Management: Business-centric UI for managing tool visibility, agent permissions, and security policies.

5. Conclusion

• Use AgentGateway if you are an infrastructure provider who needs to expose MCP-based tools to multiple external applications securely and reliably.
• Use Light-Fabric if you are building intelligent agents that need to use those tools to solve complex business problems within a governed framework.