Model Context Protocol (MCP) Enterprise Agent

An enterprise-grade, sandboxed Model Context Protocol (MCP) server designed to enable LLM agents to securely interact with internal enterprise assets—specifically relational databases (PostgreSQL) and the local directory system under absolute sandbox boundaries.

Mathematical Formulation of Sandbox Security

Let $S$ represent the sandbox root directory (absolute path), and let $P$ represent the user-supplied relative path. We define the validation mapping $V: P \to \mathbb{R}$ such that:

$$V(P) = \text{resolve}(S, P)$$

The safety condition is governed by the prefix assertion:

$$V(P) \cap S = S \quad \text{and} \quad |V(P)| \ge |S|$$

If $V(P)$ violates this constraint (e.g., through directory traversal sequence ../), the execution layer triggers an immediate transaction abort and raises an McpError with error code INVALID_PARAMS (representing a security boundary transgression).

Related MCP server: mcp-devtools

System Architecture

graph TD
    A[LLM Orchestrator / Client] -->|JSON-RPC 2.0| B[Stdio Server Transport]
    B --> C[Enterprise MCP Server]
    C -->|Security Validator V| D{Path Boundary Check}
    D -->|Passed| E[Secure Sandbox FS]
    D -->|Failed| F[Raise Security Abort]
    C -->|Read-only Enforcer| G[PostgreSQL Client]

System Requirements

• Node.js >= 20.0.0

• Docker >= 24.0.0

• PostgreSQL >= 15.0

Getting Started

Local Setup

1. Compile the server:

   npm install
   npm run build

2. Run with standard input/output transport:

   npm start

Docker

docker build -t mcp-enterprise-agent .
docker run -i --rm -e DATABASE_URL=postgresql://user:pass@host:port/db mcp-enterprise-agent

Benchmarks