A new version of the Model Context Protocol lands July 28th: stateless transport, deprecations, tasks and MCP apps.
'MCP is dead, now there are CLIs and skills.' No. MCP was never about the toys on your laptop - it's about composable AI agents used by groups of people over the network.
Under the hood, ChatGPT apps are just MCP servers, and this is why MCP is quietly winning while Google A2A struggles.
Both MCP and A2A now live under the Linux Foundation, and here is why the AI field badly needs these protocols to finally unify.
The new TOON format promises to save you tokens at scale, but you spend more tokens explaining it to the model than you ever save - and no, you are not scaling anyway.
A tool is a function the language model chooses to call; a resource is something read into context. Here is how the three MCP constructs actually differ.
A journalist copy-pasting straight from ChatGPT is not a gaffe, it is the use case handed to you on a plate: the agent should send the work where it belongs, with human review by design.
Three things worth noting: Meta's video world models, Karpathy's Software 3.0 talk, and Google donating the A2A protocol to the Linux Foundation.
