I’ve been toying with building remote MCPs on Cloudflare workers. I came in with the idea that “you could probably use REST APIs for everything” then implemented both REST, and MCPs side by side. Also built an SSE and a “streamable http” version.

For building apps that call the server, using the APIs was way easier.

For building an LLM system for figure out what API tool calls to make, it’s quite a bit of work to recreate what the MCP folks did.

I think MCPs are a huge time saver for wrapping AI around a bag of tools, without having to hard code every API call and interaction.

If anything, I think using MCPs is a massive convenience and time saver for building and prototyping LLM + tool calling apps.

Also for SSE vs Streamable HTTP, I don’t think “streamable” uses SSE at all? I think the problem they were solving for was the annoying long lived SSE connections — you can definitely see the difference on Workers though, switching away from SSE makes the workers way faster for multiple connections

Edit: the experience of building on Cloudflare and Claude is extremely frustrating; Claude is unable to output errors properly, can’t edit config inside the app, and has to be restarted constantly. Cloudflare stops working properly on SSE connections randomly, and throws bizarre errors every now and then.