The Registry-Runtime Gap in Agentic AI: A 35-Day Measurement of 31,407 Public Model Context Protocol Endpoint Records

AI agents increasingly reach databases, APIs, communication systems, payment services, and developer infrastructure through external tool servers. Yet ecosystem catalogs and repository listings describe availability, not whether an endpoint can complete a live protocol handshake, enumerate a usable tool surface, or publish machine-readable information about potentially state-changing operations. This retrospective observational preprint reports a 35-day measurement of the public Model Context Protocol ecosystem, covering 31,407 deduplicated discovered endpoint records and 152,781 repeated observations collected between 18 June and 23 July 2026. In the analyzed baseline, 4,395 endpoint records completed an anonymous MCP handshake and 4,162 exposed an enumerable tool surface. These surfaces declared 78,041 tools. A capability classifier identified 30,641 tools as potentially state-changing, of which 7,618 published none of the measured machine-readable safety annotations. The findings identify a registry-runtime gap: public discoverability cannot be treated as evidence of current operational readiness, and machine-readable disclosure remains incomplete among observable tool surfaces. Non-responsive endpoints are not necessarily defective, and missing annotations do not by themselves establish unsafe behavior. This study uses a research snapshot derived from SaSame’s continuously operating MCP observation infrastructure. The production observation system and its population continue to evolve independently of the research snapshot. Conflict of interest: The author is the founder and owner of SaSame S.R.L., which operates the measurement infrastructure used in this study and provides MCP production, observation, and repair services.