MCP Code Execution Agent Example

This project is a minimal TypeScript implementation demonstrating Anthropic's "Code Mode" or "Code execution with MCP".

Instead of overwhelming the language model's context window with dozens of tool definitions, the agent is given a single execution tool (execute_ts_code). At startup, the host application dynamically interrogates attached MCP servers (e.g., the local filesystem or a memory graph), and generates TypeScript stubs on your local disk in the .mcp-tools directory. The AI agent can then read these stubs using standard filesystem tools and write self-contained scripts to orchestrate the tools directly in a sandbox.

Prerequisites

• Node.js environment
• TypeScript and TSX
• An Anthropic API Key

Getting Started

1. Install Dependencies If passing the project folder to a fresh machine, run:

   npm install

2. Add Your API Key Create a .env file in the root of the project with your Anthropic Key, or simply export it in your terminal:

   export ANTHROPIC_API_KEY="sk-ant-api03-..."

3. Run the Demonstration

   npm run start
   # or
   npx tsx src/index.ts

Architecture

• .mcp-tools/: Contains the auto-generated TypeScript stubs. When index.ts runs it creates these physically on disk so the LLM can discover them.
• src/index.ts: The host. Connects the MCP Servers, spins up a local proxy, and then starts the agent.
• src/agent.ts: The Anthropic loop. It tells the agent about the .mcp-tools/ directory and prompts it to write self-contained scripts.
• src/executor.ts: Uses tsx to evaluate the TypeScript scripts generated by the agent.
• src/mcp-client.ts: Wraps the generic @modelcontextprotocol/sdk for easy transport management.