The Model Context Protocol got a significant overhaul this week. Version 2.0, officially called the “2026-07-28 specification,” strips out the biggest pain point that’s kept MCP from wider adoption: state management.

If you tried MCP when Anthropic launched it in November 2024, you probably remember the core problem. Every server had to maintain persistent connections and track conversation state. That meant infrastructure. That meant keeping processes alive. That meant dealing with all the edge cases that come with stateful systems: reconnections, timeouts, cleanup, the whole mess.

The new spec throws that out. MCP servers are now stateless by default. Each request is self-contained. The client sends everything the server needs to respond, the server responds, done. No persistent connections, no state to manage, no background processes to babysit.

This isn’t a minor change. It fundamentally alters what it takes to build an MCP server. You can now write one as a simple function that takes input and returns output. Deploy it as a serverless function. Run it in a container that spins up and down. Host it anywhere that can handle HTTP requests.

Simon Willison, who’s been vocal about MCP’s complexity issues, says this “recaptured [his] interest” in the protocol. He immediately built two new tools: mcp-explorer for browsing and testing MCP servers, and datasette-mcp which exposes Datasette instances as MCP servers.

What this means for developers

If you’re building LLM applications, MCP 2.0 is worth another look. The original promise was a standard way to give LLMs access to tools: databases, APIs, search engines, whatever. That promise is the same, but the implementation complexity just dropped by an order of magnitude.

You can now write an MCP server in any language that can handle HTTP and JSON. No special frameworks required. No long-running processes. Just input, processing, output.

For context: the original MCP required maintaining WebSocket connections or stdio streams. You needed to think about connection lifecycle, state synchronization, and all the infrastructure that comes with that. The new spec makes MCP servers as simple to write as REST endpoints.

What shipped alongside

Willison also released a few other tools worth noting:

smevals is a new evaluation framework he built with Jesse Vincent’s Prime Radiant lab. It’s designed for running small, focused eval suites across different models and prompt configurations. The pitch is simplicity: uvx smevals and you’re running evals. If you’ve been frustrated by heavyweight eval frameworks, this might be worth checking out.

datasette-agent 0.4a0 adds a browser_task() mechanism that lets agent tools execute JavaScript directly in the user’s browser. This is genuinely interesting. It means Datasette Agent plugins can now manipulate the browser environment, not just respond with data. The security implications are obvious, but so are the capabilities.

llm-mcp-client 0.1a0 is a new MCP client for Willison’s LLM CLI tool. It lets you use MCP servers from the command line, which is exactly the kind of thing that makes new protocols actually usable.

And one more model

DeepSeek dropped DeepSeek-V4-Flash-0731 this week. It’s 304 billion parameters with “substantially enhanced agentic capabilities,” whatever that means in practice. The interesting part: it’s punching above its weight class, ranking ahead of larger models on Artificial Analysis benchmarks.

Pricing is $0.14/million input tokens, $0.27/million output. That might make it the best intelligence-per-dollar available right now. The full model is 167GB on Hugging Face if you want to run it yourself, though you’ll need the hardware to match.

The MCP changes matter more. Stateless MCP isn’t just easier to build with—it’s the kind of simplification that might actually get the protocol used beyond early adopters. The original spec required too much infrastructure. This one might not.