Cloudflare's Matt Carey has a specific claim: most engineers are using MCP wrong. His alternative, Code Mode, lets a single MCP server expose all 2,500 Cloudflare API endpoints inside roughly 1,000 tokens of context. The mechanism is a dynamic Worker loader that executes model-written code inside a V8 isolate, keeping the footprint small and the execution safe. That is not a minor implementation detail. It is a architectural inversion of how MCP is typically deployed.
The conversation goes further than the headline trick. Carey walks through where memory actually fits in agent workflows, his personal Claude setup, and Zaggy, a git wrapper he built specifically to stop agents from force-pushing his repositories. These are not hypothetical concerns. They are the friction points that appear when you run agents against real infrastructure, and the solutions reveal assumptions worth questioning.
The full episode is worth reading for the CodeAct paper reference, the discussion of server-side versus client-side Code Mode, and Carey's position on where agent memory is heading. The 1,000-token number is the hook. The V8 isolate execution model and the Zaggy story are why you stay.
[READ ORIGINAL →]