# How reliable are MCP tool calls in production?

Not very, on average. A [2026 stress test of 100 MCP servers](https://www.digitalapplied.com/blog/mcp-server-reliability-100-server-stress-test-study) 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:
- Digital Applied, "100 MCP Servers Stress-Tested: Reliability Findings": https://www.digitalapplied.com/blog/mcp-server-reliability-100-server-stress-test-study (fetched 2026-08-25)

Source: https://tessary.ai/answers/tool-calling/how-reliable-are-mcp-tool-calls-in-production
More on Tool calling: https://tessary.ai/answers/tool-calling
From Tessary, agent reliability for AI agents in production: https://tessary.ai