Your MCP server eats 55,000 tokens before your agent says a word -- I measured the real cost

Short summary

MCP server tool definitions consume massive tokens on every conversation turn—PostgreSQL uses 35, full GitHub uses 55,000—degrading model performance even when tools aren't used. Three optimization strategies reduce overhead: filter exposed tools via allowedTools, trim verbose descriptions, and disconnect unneeded servers. The emerging MCP Tool Search protocol promises a 95% reduction in startup token cost.

• •Full GitHub MCP server costs 55,000 tokens per turn vs. 35 for PostgreSQL—a 1,500x cost difference
• •Tool schemas load on every conversation turn regardless of use, degrading model output quality past 50+ tools
• •Reduce overhead via tool filtering (270→10 tools saves 96%), description trimming, and disconnecting unused servers

Generated with AI, which can make mistakes.