MCP Client with BAML

[qdrddr]

You know how LLMs can get confused if you give them more than 20 tools in IDEs like Cursor (and any others obviously)? Well, it seems like even regular users will end up with hundreds or even thousands of tools because each MCP server might have a few dozen.

I think BAML could be a great solution to this problem. It could use the embedding model to help narrow down those tools to 20, so they can be used reliably.

https://github.com/hellovai/ai-that-works/tree/main/2025-03-31-large-scale-classification

Would you guys be so kind and provide a simple example of an MCP Client implementation using BAML to limit number of tools and select the lools reliably with BAML? 🤗

[linear]

BAML-359 MCP Client with BAML

[Elijas]

sounds good :slight_smile:
It would be should be really straighforward to do with baml-agents

1. ActionRunner.actions is a list of all the registered tools in a single unified formatf (name + description + params_json_schema).
2. Then when it's time to make a call, Create a def include(action_id: str) -> bool that returns True for all the action_id that make it to the top 20 and pass this include straight into BAML function wrapper

It would take like 10 mins to implement, I could give you the code if you'd like

But i'd love to see an official BAML approach to this as well