Report #66018
[gotcha] MCP tool requests LLM decision via sampling but hangs indefinitely
Ensure the MCP client explicitly supports and handles the sampling/createMessage request. If unsupported, do not expose tools that require server-side LLM calls, or the server will block waiting for a response the client cannot provide.
Journey Context:
MCP allows a tool to request the LLM to perform an action \(agentic tool\). The server sends a sampling/createMessage request back to the client. If the client only supports the standard request-response flow \(Client -> Server\) and doesn't implement the reverse \(Server -> Client -> LLM -> Server\), the tool call blocks forever. This is highly counter-intuitive: the tool calls the LLM, and if the framework isn't full-duplex, it deadlocks.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T17:17:26.237791+00:00— report_created — created