Reliability Index

• This is not uptime. It is the share of our checks that succeeded. A failed check can be their outage, our runner, or the network. Only the first is a fact about the project, and we only started recording which was which on 2026-09-01.
• An MCP handshake is strong evidence. An HTTP reach is weak. A parked domain answers HEAD. A tools/list response cannot be faked by a parked domain.
• “Answered a test call” is stronger than “answers a handshake”. Each MCP server is asked one read-only call a day, picked from its own tool list: a list_, get_ or search_ name carrying no mutating verb anywhere in it (get_swap_quote is refused despite the prefix), taking no required arguments and declaring none that look like an address, a key or an amount. No pill means we did not ask, never that the server failed. An error is not a verdict either: one common answer is a server telling us an argument it did not mark as required is in fact required, which is the server working.
• “Empty tool list” means the handshake succeeded and the server advertised nothing. The endpoint answered, so it counts as answered, but an agent that finds it sees nothing it can call. Servers that need a login can look like this to us.
• Before 2026-09-01 a bot-block (HTTP 403) counted as a failure. When an earlier check failed, the rate and the record length on that row cover only the checks since that date (a record with no earlier failure is counted whole), and a listing needs 14 of them to be ranked. Rows whose failures all have an unknown cause are shown but ranked last, marked cause unknown.
• Install reproduced means we ran the documented install and entry in a throwaway container, on the date of the check. It says nothing about runtime safety.
• The daily bars show the last 30 days, one bar per day. A short grey bar means no check was counted that day: no check ran, or the miss was on our side. It is a gap in our record, not a failure.
• Colours group the number, they do not grade it. 95% and up, 80% to under 95%, and under 80%. Grey means the cause of every miss is unknown.

Take the data. JSON or CSV (CC-BY), with the columns observed_success_pct, observed_days and install_verified_at. Over MCP: search_resources(min_observed_success).