Four tools beat both fine-grained primitives and one monolithic tool.

MCP-GRANITE treats tool-interface granularity as a controlled variable, running 81 multi-step scenarios across 9 domains at 4 granularity levels over 9 locally deployed models from 268M to 20.9B parameters, 8,748 trials total. A 4-tool interface improved task completion 16.4% over primitives and 33.6% over a single tool, while nearly doubling argument accuracy. Model size correlated only weakly with completion and strongly with latency, which is the finding that matters at the edge where scaling up isn't available. (arXiv 2609.24161)