A Survey on Model Context Protocol: Architecture, State-of-the-art, Challenges and Future Directions
Partha Pratim Ray
Abstract
The rapid proliferation of large-language-model (LLM) services is constrained by three long-standing barriers: stateless integration interfaces, ad-hoc security controls, and the absence of a unified mechanism for injecting rich, multi-turn context. The Model Context Protocol (MCP) has been proposed to overcome these limitations through a session-oriented, JSON-RPC framework that allows LLMs to negotiate capabilities, invoke external tools, and retrieve contextual resources under fine-grained, OAuth 2.1-compliant access control. This article provides the first comprehensive treatment of MCP from a communications-systems perspective. We (i) dissect its layered host-client-server architecture, (ii) formalize its lifecycle and transport semantics, and (iii) benchmark official and community implementations spanning software-development, cloud-automation, healthcare, and cybersecurity domains. Open challenges-including transport-layer latency,token-lifecycle overhead, cross-server privilege escalation, and persistent-context tampering-are analysed in depth, and mitigation strategies rooted in cryptographic hardening, edge caching, and capability-graph isolation are articulated. Finally, we outline a future research agenda that extends MCP toward multimodal payloads, Quick UDP Internet Connection (QUIC)-enabled streaming, and post-quantum credential exchange. By synthesizing protocol design with deployment evidence, the article positions MCP as a pivotal enabler of secure, adaptive, and scalable agentic workflows-marking a decisive forward step in the maturation of real-world LLM applications.