Code Mode for MCP Servers

What is Code Mode?

When an MCP tool returns a large API response (e.g. listing 500 Kubernetes pods, 200 SCIM users, or thousands of GitHub issues), that entire payload enters the LLM's context window — consuming tokens and degrading performance.

Code mode flips the approach: instead of dumping raw data into context, the LLM writes a small processing script. The MCP server runs the script in a sandboxed runtime against the raw data, and only the script's stdout enters context.

This works especially well with well-known APIs (SCIM, Kubernetes, GitHub, Stripe, Slack, AWS, etc.) because the LLM already knows the response schema from training data — it can write the extraction script in one shot without inspecting the data.

Typical results: 65–99% context reduction.

Inspiration