It turns out we've all been using MCP wrong. Most agents today use MCP by exposing the "tools" directly to the LLM. We tried something different: Convert the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. The results are striking.

Cloudflare's platform is a leading provider of internet security and performance solutions, offering insights into web security, content delivery, and DNS management. Through documentation, blog posts, and webinars, Cloudflare provides insights into protecting websites and applications from cyber threats and improving performance. Developers and IT professionals can learn about CDN (Content Delivery Network), DDoS mitigation, and firewall configurations to secure and accelerate web traffic.

Cloudflare

Cloudflare introduces "Code Mode" for MCP (Model Context Protocol), where instead of exposing tools directly to LLMs, the system converts MCP tools into TypeScript APIs and asks LLMs to write code that calls these APIs. This approach leverages LLMs' superior code-writing abilities compared to tool-calling, enables handling of more complex tools, and improves efficiency when chaining multiple calls. The implementation uses Cloudflare Workers' V8 isolates for secure sandboxing, with a new Worker Loader API that creates lightweight, millisecond-startup sandboxes for executing agent-generated code.

Code Mode: the better way to use MCP

Dynamic Worker loading: no containers here