What breaks in MCP servers: results from launching 750 registry entries

In the latest MCP server test, a sample of 750 registry entries was evaluated to identify potential issues and breakage points. The test, conducted on October 19, 2026, involved testing the servers in a clean sandbox environment without supplying any API keys.

The results showed that 57% of the servers were working, with a median handshake time of 563 milliseconds and a median tools/list time of 14 milliseconds for npm servers. PyPI had a median handshake time of 1,581 milliseconds and a median tools/list time of 16 milliseconds. HTTP and SSE servers had 65% and 10% working entries, respectively, with median handshake times of 436 milliseconds and 1,152 milliseconds.

The failure reasons for the non-working servers varied. 92 servers failed due to missing configuration or environment variables, while 88 were affected by dead remote endpoints. 54 servers required credentials, but these were not tested in this particular run. Additionally, 47 servers crashed on start, 12 experienced install failures, and 10 had handshake failures.

Upon reviewing the results, a practical checklist was provided for server authors. These recommendations include not crashing when an environment variable is missing, finishing the handshake and returning clear errors from tool calls, monitoring remote endpoints, pinning dependencies, and ensuring fast startup times to prevent client timeouts.

Notably, 69 out of the 100 servers utilizing SSE entries were found to be non-functional. While install failures were relatively few at 12, each one posed a significant issue for users. The trend in working counts over the past three weeks revealed a slight increase, with 431 working servers reported on October 19, 2026.