Code Execution with MCP: Building More Efficient Agents

Published on Anthropic's engineering blog on 4 November 2025 by Adam Jones and Conor Kelly, this article diagnoses two costs that appear once an agent connects to many Model Context Protocol servers: every tool definition is loaded into the context window upfront, and every intermediate result flows back through the model during multi-step work. Anthropic's answer is to stop presenting MCP tools as a flat list of callable functions and instead expose each server as a directory of typed TypeScript modules on a filesystem the agent can explore - servers/google-drive, servers/salesforce and so on, one file per tool with a typed interface. The agent then writes code that imports only the handful of tools a task actually needs, so tool discovery becomes progressive disclosure rather than a fixed context tax paid on every request. The worked example combines fetching a Google Drive document and updating a Salesforce record in a single script rather than a chain of model round-trips, and the headline figure is a scenario dropping from roughly 150,000 tokens to 2,000, which the authors present as a 98.7% time and cost saving. The piece also covers filtering large datasets inside the execution environment before anything reaches the model, tokenizing sensitive fields so private data never enters context, and persisting state to files across steps.

Topic

MCP

Level

Intermediate

Format

Guide

Cost

Free

Duration

~20 min read

Provider

Anthropic

Hands-on

No

Certificate

None

• ✓Diagnose context bloat caused by loading every MCP tool definition upfront
• ✓Present MCP servers as a filesystem tree of typed TypeScript modules
• ✓Apply progressive disclosure so the agent loads only the tools needed
• ✓Filter and aggregate large tool results inside the execution environment
• ✓Chain multi-server workflows in code instead of sequential model round-trips
• ✓Keep sensitive fields out of model context using tokenization techniques
• ✓Persist intermediate state across steps so long tasks survive interruptions

• •Quantifies the win concretely: about 150,000 tokens reduced to roughly 2,000
• •Comes from the team that authored the Model Context Protocol itself
• •Shows a real filesystem layout and typed interfaces, not abstract architecture advice
• •Triggered a wave of community implementations tracked in the MCP discussions
• •Covers privacy and state persistence, not only the token-count headline

This page was written from 3 sources, 2 on domains other than anthropic.com.

1. 1.anthropic.com — code execution with mcpvendor
2. 2.news.ycombinator.com — item
3. 3.github.com — 629