MCP failure results are observable but rarely actionable: typed fields exposed cause in zero of 21 induced failures

arXiv·medium signalXBlueskyLinkedInCopy link

A client receiving `isError:true` knows something went wrong but may have no machine-readable basis for choosing between fixing an argument, authenticating, waiting, picking another tool, or stopping. Auditing 21 safely induced failures across ten reachable sampled MCP servers with a six-part actionability profile, typed fields exposed failure in 18 cases and a broad policy in 8, yet exposed no specific cause, target, executable repair, or replay constraint in any case. Prose error text carries more cause and target information but makes semantic interpretation part of the recovery path, meaning deterministic MCP error handling is currently not possible from the result alone.