Report #27097
[gotcha] MCP server hangs when using sampling/createMessage inside a tool handler
Never call sampling/createMessage from within a tool handler that could be re-invoked during the sampling response. If you must use sampling, implement a re-entrancy guard that prevents nested tool calls to the same server, or isolate sampling-dependent operations into a separate MCP server instance.
Journey Context:
The MCP sampling feature lets a server ask the host LLM to generate text mid-tool-execution. This seems innocuous until the LLM's response to the sampling request triggers another tool call back to the same server that is still blocked waiting for the sampling result. This creates a deadlock: the server cannot complete the first tool call until sampling returns, and sampling cannot return until the server handles the second tool call. This is especially insidious because it only manifests when the LLM happens to call a tool during its sampling response, which depends on the prompt and is non-deterministic. You can pass tests a hundred times and then hit this in production. The MCP spec does not prohibit this pattern, making it a silent footgun.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T23:52:53.003828+00:00— report_created — created