How reliable are MCP tool calls in production?

Not very, on average. A 2026 stress test of 100 MCP servers found a median per-call pass rate of 71%, and reliability across servers was bimodal rather than clustered near that median: some were solid, many were not. The number that matters more than the median is what happens when calls chain. At 71% per call, five sequential tool calls succeed end to end only about 18% of the time, and ten calls drop to roughly 3%. An agent does not need one reliable call, it needs every call in its chain to land.

The same study found a separate gap that makes chains worse: 71% of bottom-decile servers have no idempotency protection. Its authors call the most common reliability problem “a successful first call that the agent mis-classified as failed and then re-invoked.” Retrying a call you cannot confirm failed risks duplicating whatever side effect it already had.

sources

keep reading

More on this.

Why did my agent call the wrong tool with invented arguments?The model matched your request to the tool whose description sounded closest, then filled a gap in the arguments with a plausible guess instead of a real value.answer →Can parallel tool calls break if one depends on the other running first?Yes. Running two tools at once removes the ordering their logic depended on, and the dependent one can execute before its precondition exists.answer →Does tool selection accuracy get worse as an agent gets more tools?Yes. Narrowing a five-tool list down to only what a query needs raised selection accuracy on medium-difficulty queries by 16 percentage points in one study.answer →Everything on tool callingThe definition, and every question answered under tool calling.concept →

Send us the traces you already emit.

book a demo