Design your MCP server like a UI, not an API (bump.sh)

Recent discussions in the AI community suggest a decline in the relevance of Model-Connected Platforms (MCP) in favor of Skills and CLI combinations. However, a new perspective argues that MCP is still essential for delivering product capabilities to end users via AI assistants. The primary challenge with MCP has been "context bloat," where servers unnecessarily inflate the context size with excessive tool definitions—leading to inefficiencies. For instance, a user measured that just seven MCP servers consumed over 67,000 tokens even before initiating a single prompt. The article provides strategies to optimize MCP server design by treating it not as an API but as a user interface intended for the end user. By focusing on user intentions and minimizing tool exposure, developers can create lightweight, efficient servers that enhance both user experience and LLM performance. Key recommendations include limiting the number of tools, using clear and concise tool names, and filtering output data to only what is absolutely necessary. These practices can significantly improve usability and reduce token consumption, ultimately driving better decision-making by AI models. The insights serve as a call to consider efficient server design as critical to MCP success in AI applications.