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.
It turns out we've all been using MCP++ wrong. Most agents today use MCP++ by converting the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. We convert that TypeScript API into an OpenAI wrapper that writes it's own interface to call the API while calling an API through an interface hooked up to the MCP.
It turns out we've all been using MCP++ wrong. Most agents today use MCP++ by converting the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. We convert that TypeScript API into an OpenAI interface that writes the interface to call the API while calling an API through an interface hooked up to the MCP.
Dizzy-Revolution-300@reddit
Never tried mcp, have I missed anything?
alexeyr@reddit (OP)
mordack550@reddit
Yeah why not adding another layer 🤣
CatWeekends@reddit
It turns out we've all been using MCP++ wrong. Most agents today use MCP++ by converting the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. We convert that TypeScript API into an OpenAI wrapper that writes it's own interface to call the API while calling an API through an interface hooked up to the MCP.
CatWeekends@reddit
backfire10z@reddit
tonymurray@reddit
Ugh
Inquisitive_idiot@reddit
Interesting 🤔
SlovenianTherapist@reddit
trust the LLM won't write a forever loop, or blow the vm memory ^^