Report #91940
[frontier] MCP server needs to reason about its own output before returning result to agent
Use MCP Sampling: implement the sampling handler in your MCP server to request LLM completions from the host model. Pass the tool's intermediate result as a user message with a system prompt instructing classification, ranking, or summarization. Return the LLM-refined output to the calling agent instead of raw data.
Journey Context:
Developers treat MCP as a strict request-response tool protocol where the server returns raw data and the agent interprets it. The MCP spec defines a Sampling capability allowing servers to request LLM completions from the host. This is essential when a tool produces data requiring interpretation: a code search tool finding 20 matches can use sampling to ask the host LLM to rank them by relevance, returning only the top 3 with rationale. Without sampling, the agent wastes turns and context window space on classification the tool could have done. Tradeoff: each sampling call adds one LLM inference in latency and cost. Use when returning raw data would cost more in wasted agent turns than the sampling call costs. Never use for simple lookups or deterministic transformations.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T12:54:42.581273+00:00— report_created — created