What's the Best Practice to use the MCP Tools?

[Abel2333]

I found that in the example of rmcp official repo, they convert the mcp tools to the dynamic tools of rig with an embedding model. But in the example of rig-core/rmcp.rs, method rmcp_tools provides another way without embedding model.

My question is, what's the differents between these two way? Is there a difference in their efficiency? And why the first way need an embedding model?

[joshua-mo-143]

Hi, the reason why you might need an embedding model is for tool ragging (essentially being able to carry out semantic search and fetch tools from a vector store based on a query).

The reason why we provide another way by default is so that if you want to just add tools without tool ragging, it should be simple to do.

[Abel2333]

Thanks for your answer.

Then what's the differents between the dynamic tools and static tools? I want to create a unified interface to register local tools and MCP tools, but these different tools have almost entirely different registration methods, which leaves me confused.

[joshua-mo-143]

Thanks for your answer.

Then what's the differents between the dynamic tools and static tools? I want to create a unified interface to register local tools and MCP tools, but these different tools have almost entirely different registration methods, which leaves me confused.

Dynamic tools are selected using similarity search from a vector store, static tools will always be in the context.

If you want to use MCP tools pulled from the rmcp integration, you need to create a ToolSet from them (likely via ToolSet::from_tools()), then add them in using AgentBuilder::dynamic_tools().

The agent tool-building stuff could probably use some love in terms of the documentation/having helper methods to do this, though.

[Abel2333]

Thanks for your answer, i have the idea about how to use these tools.