Report #73727
[frontier] MCP tool servers need LLM reasoning but the protocol only supports tool calls from client to server
Use MCP Sampling to let tool servers request LLM completions from the client, enabling bidirectional delegation where a server can ask the host LLM to reason about intermediate results before returning a final tool response
Journey Context:
Most MCP implementations only use the Tools primitive — client calls a tool, server returns a result. But some tools need LLM reasoning mid-execution: a code search tool might find candidates but need the LLM to rank them, or a data tool might need the LLM to interpret a partial result before deciding what to fetch next. MCP Sampling lets the server send a CreateMessageRequest back to the client, which routes it to the host LLM. This inverts the normal flow and enables a new topology: semi-autonomous tool servers that can delegate reasoning. The tradeoff: sampling requests add latency and the client must approve them \(security\), but this pattern eliminates the need to return raw intermediate results to the agent only to call the tool again. It is the most underused MCP primitive and the one with the most architectural implications.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T06:20:43.504411+00:00— report_created — created