The Model Context Protocol (MCP) is an open standard designed to connect AI agents with external systems, streamlining the integration process and reducing redundancy. Since its launch in November 2024, MCP has gained rapid adoption, leading to the creation of thousands of servers and SDKs across major programming languages. Developers can now build agents that access numerous tools through MCP, but as the number of integrations increases, so do challenges related to efficiency and costs. Loading extensive tool definitions and processing intermediate results can lead to high token usage and slower response times. The article highlights two main issues with traditional MCP implementations: the overload of the context window with tool definitions and the additional tokens consumed by intermediate results. For instance, when an agent calls multiple tools, it often passes large amounts of data back and forth, resulting in excessive token consumption. This can be particularly problematic with lengthy documents, where agents may exceed context limits or introduce errors when copying data. A proposed solution is to utilize code execution environments, allowing agents to interact with MCP servers as code APIs instead of direct tool calls. This method enables agents to load only the necessary tools for each task, drastically reducing token usage from around 150,000 tokens to just 2,000 tokens in certain scenarios. By using a filesystem structure for tools, agents can discover and utilize only what's needed, leading to significant efficiency gains. Code execution also allows for better handling of large datasets, filtering results in the execution environment rather than through the model, which conserves context and improves performance.