Report #50267
[frontier] MCP servers only call tools, they can't initiate LLM reasoning
Use MCP's sampling capability to let servers request LLM completions from the client via sampling/createMessage, enabling server-initiated agentic loops where the server reasons about data before returning results—eliminating multi-turn orchestrator round-trips.
Journey Context:
Most MCP implementations use only tools and resources. The sampling primitive lets a server send a createMessage request to the client, which forwards it to the LLM and returns the response. This enables a code-analysis server to validate its own outputs via LLM before returning, or a data server to summarize and reason about query results internally. The tradeoff is increased latency and cost per tool call, and the client must explicitly grant sampling permissions. People commonly treat MCP as purely request-response tool calling, missing that the protocol was designed for bidirectional agentic communication. The alternative—forcing the orchestrating agent to make multiple round-trips for each sub-reasoning step—wastes context window space and adds orchestration complexity. Sampling keeps sub-reasoning localized to the server that owns the domain.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T14:51:29.539680+00:00— report_created — created