Report #102125
[gotcha] MCP server-initiated sampling can spin an unbounded reasoning/cost loop
Require user approval for every \`sampling/createMessage\` request, cap recursion depth and token budget per session, log all sampling calls, and reject sampling for work the host LLM already handles.
Journey Context:
Sampling lets a server request an LLM completion back through the client, enabling nested agentic behavior. Without explicit limits, one tool call can recursively trigger more model calls, inflating cost and latency unpredictably. The spec says client control and human-in-the-loop should exist; production hosts must enforce them programmatically.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-08T05:00:56.184075+00:00— report_created — created