Tool Attention Reduces MCP Token Overhead in LLM Agent Workflows

A recent study published on arXiv (2604.21816) presents Tool Attention, a middleware solution aimed at addressing the 'MCP Tax' or 'Tools Tax' in scalable agentic workflows. The Model Context Protocol (MCP) typically links LLM agents with external tools but results in a per-turn overhead of 10k to 60k tokens due to its stateless and eager schema injection. This excess burden enlarges the key-value cache and hampers reasoning when context usage hits 70%. Tool Attention extends the 'Attention Is All You Need' framework, shifting from self-attention on tokens to gated attention for tools. It integrates an Intent Schema Overlap (ISO) score derived from sentence embeddings, a state-aware gating mechanism for preconditions and access scopes, along with a two-phase lazy schema loader, aiming to lower recurring token budget costs.