AgentMesh-MCP: A Secure and Governed Framework for Agentic AI Systems Using LLM Agents and Model Context Protocol Servers

Agentic AI systems will shape software engineering, cloud operations, enterprise automation, research, education, cybersecurity, legal work, healthcare administration, and many other domains. Their value comes from the ability to connect reasoning with action. Their risk comes from the same source. LLM agents that can use MCP servers are more interoperable and more capable, but they also create new failure modes involving prompt injection, tool poisoning, malicious servers, credential misuse, data leakage, excessive agency, memory poisoning, and multi-agent compromise. This paper proposed AgentMesh-MCP, a secure and governed framework for LLM agents and Model Context Protocol servers. The framework separates reasoning from execution, routes MCP calls through a policy and authorization gateway, filters context, scores risk, requires human approval for high-risk actions, monitors tool execution, records immutable audit events, supports rollback and recovery, and provides governance dashboards. It also formalized the system as a constrained transition process and supplied algorithms, evaluation metrics, experimental design, findings, use cases, limitations, and future research directions. The main conclusion is that production-grade agentic AI should be designed as governed distributed infrastructure, not as a more elaborate chatbot. MCP can become a powerful interoperability layer for AI systems, but secure deployment requires policy, identity, observability, and human control around the protocol. AgentMesh-MCP offers one path toward that goal. Its broader significance is that it treats autonomy as an engineering variable to be measured, constrained, and governed. If implemented carefully, secure MCP-mediated agents can transform digital work while preserving accountability and institutional control

[1] Anthropic, "Introducing the Model Context Protocol," Nov. 2024. [Online]. Available: https://www.anthropic.com/news/model-context-protocol [2] Model Context Protocol, "What is the Model Context Protocol (MCP)?" 2026. [Online]. Available: https://modelcontextprotocol.io/docs/getting-started/intro [3] Model Context Protocol, "Architecture overview," 2026. [Online]. Available: https://modelcontextprotocol.io/docs/learn/architecture [4] JSON-RPC Working Group, "JSON-RPC 2.0 Specification," 2013. [Online]. Available: https://www.jsonrpc.org/specification [5] S. Yao et al., "ReAct: Synergizing reasoning and acting in language models," arXiv:2210.03629, 2022. [6] T. Schick et al., "Toolformer: Language models can teach themselves to use tools," arXiv:2302.04761, 2023. [7] N. Shinn et al., "Reflexion: Language agents with verbal reinforcement learning," arXiv:2303.11366, 2023. [8] G. Wang et al., "Voyager: An open-ended embodied agent with large language models," arXiv:2305.16291, 2023. [9] Q. Wu et al., "AutoGen: Enabling next-gen LLM applications via multi-agent conversation framework," arXiv:2308.08155, 2023. [10] J. S. Park et al., "Generative agents: Interactive simulacra of human behavior," in Proc. UIST, 2023.