Based on posts like this one, tool use is very unnatural for LLMs. MCP is a great way to standardize protocols, but LLMs seem to do better when writing traditional code that they had training examples for.
It's possible this could be mitigated e.g. with adaptation layers on the tools themselves (#15), but it's worth exploring other alternatives that are model-agnostic.
Based on posts like this one, tool use is very unnatural for LLMs. MCP is a great way to standardize protocols, but LLMs seem to do better when writing traditional code that they had training examples for.
It's possible this could be mitigated e.g. with adaptation layers on the tools themselves (#15), but it's worth exploring other alternatives that are model-agnostic.