Anthropic just published an article on code execution with MCP, and unless I’m reading this wrong, they’re saying something pretty interesting:

1. Don’t use MCPs to call the tools directly—they’re too heavy
2. Use the MCPs to figure out how the tool works
3. Then have your agents write actual code that calls the API efficiently

If that’s what they’re proposing, it’s a clever approach to the token efficiency problem.

Instead of exposing thousands of tools as function calls (which consumes massive context), you use MCP servers as documentation that agents read, then write efficient code to interact with those APIs directly.

They claim this reduced token usage from 150,000 to 2,000 tokens—a 98.7% reduction.

Worth reading if you’re building with agents and MCPs.