Cloudflare's Matt Carey has a specific claim: most developers are using MCP wrong. In a conversation with the Changelog, Carey, who works on the Agents SDK and MCP at Cloudflare, explains how server-side Code Mode compresses all 2,500 Cloudflare API endpoints into roughly 1,000 tokens of context. The mechanism is a dynamic Worker loader that executes model-written code inside a V8 isolate, letting the model write code against an API rather than consuming hundreds of discrete tool definitions.
The technical argument is worth understanding in full. Traditional MCP implementations expose each API action as a separate tool, which burns context fast and scales poorly. Code Mode inverts this: the server exposes a code execution primitive, and the model figures out how to call the API itself. Carey also built Zaggy, a git wrapper that prevents agents from force-pushing his repositories, a detail that points at real production friction with autonomous coding agents.
The interview goes further than the architecture. Carey describes his own daily workflow with Claude, where he thinks memory fits into the agent stack, and why the current tooling around agent persistence is still unsolved. If you are building with MCP or evaluating agentic infrastructure, the Code Mode posts on the Cloudflare blog and the GitHub repo for the Cloudflare MCP server are the concrete starting points.
[READ ORIGINAL →]