Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)

The Model Context Protocol has undergone its most significant update since its initial launch in late 2024 with the release of the stateless MCP specification, also known as MCP 2.0. The previous stateful version required multiple HTTP requests to initialize a session and obtain a session ID before calling a tool. The new stateless specification simplifies this process by allowing everything to happen in a single HTTP request, which greatly reduces implementation complexity for both clients and servers and fits better with scalable web architectures.

This technical shift has revived interest in MCP, particularly because giving AI agents open shell and internet access can be risky and demands powerful models. In contrast, MCP tools are easier to audit, control, and operate reliably even with smaller laptop-based models. To explore the new specification, three new Python-based projects were developed: mcp-explorer as a command-line tool for probing MCP servers, datasette-mcp as a plugin adding an endpoint to Datasette instances for running read-only SQL queries, and llm-mcp-client for integrating MCP capabilities directly into the LLM tool.

Ultimately, the stateless update makes MCP a more practical and secure standard for exposing tools to AI agents compared to open shell environments.