AgentMesh-MCP: A Secure and Governed Framework for Agentic AI Systems Using LLM Agents and Model Context Protocol Servers
Year: 2026
Type: article
Abstract: This paper proposes AgentMesh-MCP, a secure and governed framework for agentic artificial intelligence systems that combine large language model agents with Model Context Protocol servers. Agentic AI systems increasingly reason over goals, decompose tasks, call tools, query private data, invoke APIs, and coordinate with other agents. The Model Context Protocol provides a standardized client-server mechanism for exposing tools, resources, and prompts to AI applications, but protocol-level interoperability alone does not solve authorization, prompt injection, tool poisoning, auditability, or unsafe autonomy. We formalize the agentic system as a constrained partially observable decision process in which agent state, environment state, tool state, policy state, and governance constraints evolve through MCP-mediated actions. The proposed framework introduces a policy and authorization gateway, context firewall, risk scoring engine, human approval gateway, execution monitor, immutable audit ledger, red-team module, evaluation engine, rollback layer, and governance dashboard. We synthesize evidence from LLM agent research, tool-use systems, prompt-injection studies, AI risk-management standards, and official MCP documentation to derive architectural requirements and evaluation criteria. The paper contributes a threat model, formal model, algorithms for policy-gated tool invocation and secure context filtering, comparative analysis, and a reproducible experimental testbed. The central finding is that production-grade agentic AI should be treated less as a chatbot extension and more as a distributed socio-technical system whose safety depends on least privilege, context isolation, observable execution, explicit human control points, and continuous governance.
DOI: 10.55041/ijsrem62689
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7160396918