Code execution with MCP: Building more efficient agents (www.anthropic.com)

The Model Context Protocol (MCP) — an open standard for connecting LLM agents to external tools and data — has seen rapid adoption since its November 2024 launch, spawning thousands of MCP servers and cross-language SDKs. As agents scale to hundreds or thousands of tools, the common practice of loading every tool definition and passing raw intermediate results through the model inflates context windows, raises latency and cost, and can exceed token limits. The blog proposes treating MCP servers as code APIs in a code-execution environment (e.g., a TypeScript file tree where each tool is a callable module) so agents load only the interfaces they need, run filtering and transformations off-model, and then return compact, relevant summaries to the LLM. Technically, this “Code Mode” approach reduces token consumption dramatically (the post cites an example drop from ~150k to ~2k tokens, ~98.7% savings). It enables on-demand discovery (filesystem listing or search_tools with adjustable detail levels), in-execution filtering/aggregation (so only a few rows or summaries reach the model), richer control flow (loops, retries, error handling executed in code), state persistence (writing workspace files and reusable skills), and privacy-preserving workflows via client-side tokenization/untokenization of PII. The tradeoffs are added engineering complexity and runtime security considerations, but for large-scale, tool-rich agents this pattern offers a clear path to faster, cheaper, and more secure integrations.