Type: url
Original: https://www.anthropic.com/engineering/code-execution-with-mcp
Status: partial
Summary
An Anthropic engineering post (Nov 2025) arguing that agents connected to many MCP servers waste context by loading every tool definition upfront and passing intermediate results through the model, and that presenting MCP servers as code APIs the agent calls from a code execution environment addresses both. It describes a filesystem-of-tools implementation, reports a reduction from 150,000 to 2,000 tokens in one example, and lists benefits (progressive disclosure, context-efficient results, control flow, privacy-preserving operations, state persistence and skills) along with the sandboxing overhead code execution adds.