[Feature Request] Dynamic tool loading into context like Claude Desktop

[moltrus]

Can it be optimized to have the tools loaded dynamically into context using something like tool search, like in Claude Desktop? It enables support to have 100+ tools loaded into memory as it discovers them when the need arrives.

[kevinwatt]

Thanks for the request!

Claude Desktop's approach uses a tool_search step (backed by a fast model like Haiku) to semantically retrieve relevant tools — essentially RAG for tools. The tradeoff is extra latency per turn. You reduce the MCP Token but increase the round trip Token.

We're exploring alternatives: tool skills (pre-written skill files loaded based on relevance) and predictive loading (pre-fetching tools likely needed in the next turn). Both avoid the extra round-trip. Feel free to share your specific use case.