mcp-agent-enterprise-guard

Compute trust scores, permission decisions, and silent-failure risk for AI agents — pure local computation, zero API cost.

Built for founders and teams who need lightweight AI governance (especially those validating products with Prevalid).

Why this exists

AI agents often fail silently: they return empty results, pick the wrong tool, or violate simple policies without raising an error. Enterprise-style workflows (Salesforce, ServiceNow, internal data) make this painful and expensive.

This MCP server gives your agent (or your ops team) four practical tools that run entirely on the data you already have — no external APIs, no monthly data bills.

Tools

Tool	What it does
`score_agent_reliability_tool`	0–100 reliability score + breakdown + recommendations
`detect_silent_failures_tool`	Finds empty/null responses that did not set `is_error`
`evaluate_permission_tool`	Simple allow/deny rule engine for agent actions
`generate_audit_report_tool`	Human-readable markdown governance report

Quick start (local)

git clone https://github.com/princeruhulofficial/mcp-agent-enterprise-guard.git
cd mcp-agent-enterprise-guard
pip install -e .
python -m agent_enterprise_guard.server

Or with FastMCP / Claude Desktop config:

{
  "mcpServers": {
    "agent-enterprise-guard": {
      "command": "python",
      "args": ["-m", "agent_enterprise_guard.server"]
    }
  }
}

Example usage

Feed your agent run logs (list of dicts with tool_name, success, latency_ms, is_error, content) and get a reliability score + silent-failure list in seconds.

Monetization (MCPize)

Freemium model planned:

Free: limited daily calls for testing
Pro / Team / Enterprise tiers with higher limits and report export

License

MIT

Made for the Prevalid community — build AI systems people can actually trust.

mcp-agent-enterprise-guard