There is a lot of (justified) of excitement about MCP, but just like most function calling approaches, I think it has a severe limitation at the moment, and I am kind of of flabbergasted why I am the only one that is bothered by it. The problem is that when you define a tool using MCP [0] you can only specify the *inputSchema* but not the *outputSchema* (the same is the case for OpenAI [1], Claude [2], Mistral [3], Gemini [4], ...). And the optional description of a tool is just a "hint" to the model. That means that when the model needs to decide to call a tool, it can only rely on the structure of the the inputs (not even on the description). But since it does not know the structure of the output of the call, it cannot reason about the composition of multiple tool calls to solve a complex problem. Instead it needs to take a guess, call some first tool, and then interpret the output before it can reason about how to continue. The fact that tools are dynamically typed wrt to their results makes it also impossible for other tools, and humans, to reason about tools (that is a lot of tools :-). Imagine I would explain MCP to you by only showing the types you send to an MCP server, but none of the types that you receive from a server. Since tools are for all intents and purposes really just fancy REST apis, just like OpenAPI definitions [5], tool calling and MCP tools should also specify their outputs.