Model Context Protocol works for tools. It breaks for agents (blog.vtemian.com)

A new update from Claude Code introduces the Model Context Protocol (MCP), enabling server processes to interact with tools via a JSON-RPC 2.0 interface. This protocol allows plugins to communicate in a straightforward manner without the complexities of HTTP or WebSockets, providing a clean architecture where each plugin runs as an isolated child process. A key feature is the ability to convert ephemeral Claude Code sessions into permanent shareable URLs through the Claudebin tool. However, while MCP excels for simple utility tools, it falls short in agent-based systems, lacking features like lifecycle hooks and shared state that are vital for complex workflows. The significance of MCP lies in its clear separation of functionalities for tools versus agent systems. The design fosters easy portability and cross-language support, making it ideal for straightforward applications. However, developers are limited to basic interactions without the observability and flexibility needed for more dynamic applications. This shortcoming emphasizes a crucial gap in available platforms, suggesting that an optimal solution would integrate the strengths of both MCP and direct plugin models for comprehensive tool and agent orchestration. Such differentiation highlights an evolving discourse within the AI/ML community around adapting frameworks to better support diverse application needs.