mcpindex Drift Report - Edition v1 dataset

Aggregate statistics and per-server summaries of silent tool-contract changes observed across the reachable remote population of the official Model Context Protocol (MCP) registry, from mcpindex.ai's daily crawl. Contents: aggregates.json (deduped safety-relevant contract-change incidents by kind; the version-evidence split, where "silent" means the server's declared version was unchanged across the change and "changed" means it differed, with no direction asserted; annotation-flip segmentation; tool-removal counts with toolset-replacement scope; the unstable-tool class, excluded from headline counts and reported separately; coverage statistics) and per-server-summaries.csv (one row per named server with at least one safety-relevant incident). Row-level crawl history is not included. Method summary: deterministic contract diff between consecutive daily snapshots, scoped to servers reachable in both; registry entries sharing one name across multiple endpoints are merged one-record-per-name. The observation window is stated as snapshots-across-days and includes one disclosed 21-day crawler outage. Every recorded event in the underlying corpus was independently re-derived from the raw before/after contracts (a full-population census, not a sample): 0 classification errors; one pipeline defect (day-one reachability-flap artifacts) was found, excluded, and disclosed. A contract diff is not a safety verdict, and "silent" is a claim about version evidence only, not intent. Full methodology: https://mcpindex.ai/methodology The live, continuously updated counterpart of these numbers is published at https://mcpindex.ai/drift-report