The Model Context Protocol (MCP) is an open standard that enables AI agents to connect with external systems. Connecting agents to tools and data has traditionally required building a separate custom integration for every combination, resulting in fragmentation, redundant work, and significant barriers to scaling truly interconnected systems. MCP offers a universal protocol—developers only need to implement it once in their agent to gain access to an entire ecosystem of integrations. Since its launch in November 2024, adoption has been swift: the community has created thousands of MCP servers, SDKs are now available for all major languages, and MCP has become the industry’s de-facto standard for connecting agents to tools and data. Developers now routinely build agents that can access hundreds or even thousands of tools spanning dozens of MCP servers. However, as the number of connected tools increases, loading all tool definitions in advance and passing intermediate results via the context window causes agents to slow down and drives up costs. In this blog post, we’ll explore how code execution allows agents to interact with MCP servers far more efficiently—supporting more tools while consuming fewer tokens. Relying on tools consumes excessive tokens and makes agents less efficient. As MCP adoption grows, two common patterns tend to drive up agent costs and latency. Tool definitions consume excessive space in the context window. Intermediate tool outputs use up extra tokens. 1. Tool definitions consume excessive context window space.
Anthropic Engineering