Sampling Primitive: The Server Asks the Client's LLM to Generate

~12 min read

The client side always has an LLM — Sampling lets the server ask that LLM to generate completions mid-workflow, while the client still controls permissions and safety.

Sampling Primitive: The Server Asks the Client's LLM to Generate is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Sampling is server-initiated: unlike Tools/Resources/Prompts (server offers TO client), the server REQUESTS something FROM the client here
  • It works because the client side always has an LLM available — the server borrows it rather than needing its own
  • The client still controls permissions and safety — the server requests, it doesn't command
  • The book's example: an MCP server with travel tools asks the client's LLM to pick the optimal flight from a list
  • This lets a server stay 'dumb' (no embedded reasoning) while still tapping into sophisticated LLM judgment when needed