PROTO

McpVanguard

security proxyMCP enforcementdeterministic guardrails

McpVanguard is a deterministic security proxy for the Model Context Protocol (MCP). It interposes between AI agents and tool servers to enforce non-bypassable security policies with low-latency enforcement.

THE PROXY REQUIREMENT

Direct connections between agents and tools are inherently dangerous. The MCP standard does not specify authentication, authorization, or audit — leaving every Host application responsible for its own security. McpVanguard provides an execution boundary for routed MCP activity, inspecting configured tool discovery, parameter, and result surfaces before they reach the next step.

TECHNICAL ARCHITECTURE

Built in Python for rapid integration and operator customization, McpVanguard implements a deterministic rules engine that evaluates tool calls against versioned YAML manifests. It is designed to operate as a transparent proxy in front of existing MCP servers and models.

Rule Engine

Versioned YAML manifests (jailbreak.yaml, network.yaml, privilege.yaml) define deterministic matching rules. No probabilistic model judgment is involved in enforcement.

Interception Points

Tool discovery metadata checks, parameter inspection (allowlist/blocklist), selected result scanning, and execution logging (telemetry generation).

Deployment Model

Runs as a local sidecar or stdio wrapper. SSE/HTTP gateway mode is available for hosted deployments (requires API key or JWT auth). Container deployment is on the roadmap.

Performance

Low-latency deterministic enforcement. Python's flexibility enables rapid rule iteration without recompilation; production rule evaluation is optimized for throughput.

ProvnAI Mitigation

McpVanguard centralizes enforcement at the routed MCP boundary, reducing reliance on model-level self-correction or disparate server-side validations. Security events can be logged locally with optional offload to external evidence systems for compliance and incident response workflows.