Thanks! I haven't tried Rube yet, but I'm familiar with it. It's interesting for consolidating integrations, but it solves a different problem than what I'm describing here.
Rube is about "which MCP servers to use" (aggregation). This article is about "how to use MCP servers without bloating context" (efficiency).
In fact, with 500+ tools available, Rube could potentially make the token bloat problem worse unless you're selective about which tools you expose. The on-demand approach I describe would work with Rube too - you'd just call it when needed rather than loading all those tool schemas upfront.
Have you noticed context issues with Rube? Curious how they handle the tool schema size with that many integrations.
I asked about Rube, because it works in a similar way. What Rube does really well is it uses the RUBE_SEARCH_TOOLS function, and only loads the specific tools needed for your current task. It basically routes queries to the right app servers without overwhelming the context window. Nothing is frontloaded. You also don't need multiple MCPs, everything is consolidated. And on top of that, it caches the tool discoveries for 15 min, so it limits repeated searches for the same capabilities.
Great read Lakshmi. I've been using Rube for all my MCPs, have you tried it?
Thanks! I haven't tried Rube yet, but I'm familiar with it. It's interesting for consolidating integrations, but it solves a different problem than what I'm describing here.
Rube is about "which MCP servers to use" (aggregation). This article is about "how to use MCP servers without bloating context" (efficiency).
In fact, with 500+ tools available, Rube could potentially make the token bloat problem worse unless you're selective about which tools you expose. The on-demand approach I describe would work with Rube too - you'd just call it when needed rather than loading all those tool schemas upfront.
Have you noticed context issues with Rube? Curious how they handle the tool schema size with that many integrations.
I asked about Rube, because it works in a similar way. What Rube does really well is it uses the RUBE_SEARCH_TOOLS function, and only loads the specific tools needed for your current task. It basically routes queries to the right app servers without overwhelming the context window. Nothing is frontloaded. You also don't need multiple MCPs, everything is consolidated. And on top of that, it caches the tool discoveries for 15 min, so it limits repeated searches for the same capabilities.
Let me know if you ever try it :)