I Measured MCP vs Direct API Calls: The Token Math No One Tells You

Short summary

MCP agents carry 4x-32x token overhead compared to direct API calls because tool schemas are injected into every LLM message—one example burned 42,661 tokens for a simple task that needed 1,365 tokens via direct API. The overhead buys flexibility for conversational agents and centralized governance, but deterministic batch pipelines should use direct APIs instead. Field projection (returning only essential fields) can significantly reduce schema overhead.

• •MCP adds 4x-32x token overhead due to tool schema injection into every message
• •Justified for conversational agents and governance; wasteful for deterministic batch jobs
• •Field projection (returning only needed fields) optimizes schema bloat in MCP responses

Generated with AI, which can make mistakes.