What an MCP server costs you in tokens (workos.com)

A recent analysis highlights the significant token costs associated with running a Multi-Channel Processing (MCP) server, particularly focusing on tool definitions and their impact on performance and accuracy. Typically, a multiserver setup, including tools like GitHub and Slack, can consume around 55,000 tokens in tool definitions before any user interaction occurs. This high token consumption isn't just a financial burden; it also leads to diminishing accuracy in tool selection as the number of tools exceeds thirty to fifty. The recurrent bill for tool definitions in every conversation turn means that users experience slower response times and less reliable tool integration. To mitigate these issues, the article suggests several strategies, including reducing the number of exposed tools and employing deferred loading techniques to load only the necessary tool definitions when needed. This can dramatically cut definition overhead, reportedly by over 85%. Other solutions involve redesigning tool interfaces for better efficiency, allowing code execution within a sandbox environment to eliminate unnecessary data transit, and ensuring effective caching of tool catalogs to enhance performance. These insights reframe development considerations for MCP servers, emphasizing the importance of creating a streamlined and user-friendly toolset, as the tool count directly affects both operational costs and the accuracy of AI interactions.