We Cut Our MCP Token Spend in Half. Here's the Architecture

When we started scaling our MCP workflows, token usage was something we barely tracked. The system worked well, responses were accurate, and adding more tools felt like the right next step. Over time, the cost began rising in ways that did not align with how much the system was actually used. At first, we assumed this was due to higher usage or more complex queries. The data showed something else. Even simple requests were using more tokens than expected. This led us to ask a basic question. What exactly are we sending to the LLM on every call? A closer look made things clearer. The issue came from how the system was built. We handled context, tool definitions, and execution flow by adding extra tokens at every step. This article explains how we found the root cause and redesigned the architecture to fix it. The changes cut our MCP token usage by nearly half and gave us better control over how the system behaves. Understanding Token Usage in MCP Systems Once we started examining token