MCP 2.0 is live with the 2026-07-28 specification. It’s the biggest change to the Model Context Protocol since Anthropic launched it in November 2024. The central shift is moving to stateless: a single HTTP request to call a tool, instead of two. Before, you’d initialize the session and get a session ID, then call the tool. Now a single request is enough, with the method and tool name in the headers. No state to maintain on the server.

Why this matters to you. If you connect agents to external tools, MCP 2.0 solves two problems that used to be painful. MCP tools are easier to audit than giving an agent a terminal with curl, and small models running on your laptop can use them. Stateless removes session management, which was where the architecture broke when you tried to scale beyond the prototype. As we covered in the lesson on connecting your agent to your systems, MCP is the standard bridge between the agent and your data. The new version makes it viable in production.

If you want to try it: the first step is point a stateless client at a public MCP server and inspect the list of available tools. mcp-explorer and llm-mcp-client are the natural candidates to start with.

In detail

MCP was created as a standard to expose tools to AI agents uniformly. Anthropic presented it in November 2024, it had a spike of interest through much of 2025, and then was somewhat overshadowed by Skills, the approach where an agent with terminal access and internet connectivity could do much of what MCP did, more flexibly.

What changes with stateless. Stateful MCP required two HTTP requests per call: first to initialize the session and get an Mcp-Session-Id, second to execute the tool. The server had to maintain the state of each active session and route subsequent requests to the same backend machine.

The new stateless MCP uses a single HTTP request. The method and tool name travel in the headers (Mcp-Method, Mcp-Name), parameters in the JSON body, and client information in a _meta field. No session ID, no state on the server.

From an infrastructure perspective, this means you can put an MCP server behind a load balancer without worrying about session affinity. Each request is independent and can hit any backend. For those building web applications with agents, this was where the previous architecture would break down.

Tools released this week. In the days of the release, three projects came out to validate the specification in the field. mcp-explorer is a stateless Python CLI for exploring MCP servers: it lists tools, inspects their input and output schemas, and calls them by passing arguments. datasette-mcp is a Datasette plugin that exposes an MCP endpoint on any instance, with three tools: list databases, read their schema, execute read-only SQL. llm-mcp-client registers MCP tools as available tools within the llm framework.

Interest in MCP is growing again after being sidelined through much of 2025. The reason is practical: giving an agent a shell environment with internet access is risky and requires a strong model. MCP tools are simpler to control, and work even with small models running locally.

Limitations. The specification is fresh (July 28, 2026) and the ecosystem of stateless clients and servers is still thin. The three tools released this week are early releases (0.1a0). The shift from stateful MCP to stateless MCP is a breaking change, not backward compatible: anyone with legacy MCP servers needs to update them. Anyone building agents in production should evaluate migrating but shouldn’t expect the ecosystem to be mature yet.